]> Deutsches Forschungsnetz | German National Research and Education Network

Alexanderplatz 1 Berlin 10178 Germany rieckers@dfn.de www.dfn.de

Fondation Restena | Restena Foundation

2, avenue de l'Université Esch-sur-Alzette 4365 Luxembourg stefan.winter@restena.lu www.restena.lu

Security RADIUS EXTensions RADIUS TLS This document defines transport profiles for running RADIUS over Transport Layer Security (TLS) and Datagram Transport Layer Security (DTLS), allowing the secure and reliable transport of RADIUS messages. RADIUS/TLS and RADIUS/DTLS are collectively referred to as RadSec. About This Document Status information for this document may be found at . Discussion of this document takes place on the RADIUS EXTensions Working Group mailing list (), which is archived at . Subscribe at .

Introduction This document defines transport profiles for running RADIUS over Transport Layer Security (TLS) , over TCP and Datagram Transport Layer Security (DTLS) , over UDP , allowing secure and reliable transport of RADIUS messages. RADIUS/TLS and RADIUS/DTLS are collectively referred to as RadSec. The RADIUS protocol is a widely deployed authentication, authorization and accounting solution defined in , , and others. Deployment experience has shown several shortcomings, such as its dependency on the unreliable transport protocol, UDP, and its lack of confidentiality for large parts of RADIUS messages. Additionally, the confidentiality and integrity mechanisms in RADIUS rely on the MD5 algorithm , which does not meet modern security expectations. Although RadSec does not remove the MD5-based mechanisms, it adds confidentiality and integrity protection through the TLS layer. For an updated version of RadSec without need for MD5 see .

Conventions and Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here. The following terminology is used in this document:

RadSec:

A collective term for RADIUS/TLS and RADIUS/DTLS.

RADIUS/TLS:

A RADIUS exchange transmitted using TLS over TCP.

RADIUS/DTLS:

A RADIUS exchange transmitted using DTLS over UDP.

RADIUS/UDP:

RADIUS transported over UDP as defined in .

RADIUS packet:

As defined in .

RADIUS packet type:

As defined in .

(D)TLS handshake message:

As defined in TLS and DTLS .

TLS record:

As defined in TLS .

DTLS record:

As defined in DTLS . A DTLS record is always contained in one UDP datagram.

(D)TLS connection:

A single (D)TLS communication channel (with DTLS this is a synonym for Association).

UDP datagram:

A UDP packet, including the header and data.

UDP (datagram) data:

The data payload of a UDP datagram.

RadSec client:

A RadSec instance that initiates a new connection.

RadSec server:

A RadSec instance that listens on a RADIUS-over-(D)TLS port and accepts new connections.

RadSec endpoint:

A RADIUS-over-(D)TLS client or server

RadSec peer:

A RadSec endpoint connected to the RadSec endpoint that is the primary subject of discussion.

Whenever "(D)TLS", "RADIUS/(D)TLS" or "RadSec" is mentioned, the specification applies for both RADIUS/TLS and RADIUS/DTLS. Where "TLS" or "RADIUS/TLS" is mentioned, the specification applies only for RADIUS/TLS, where "DTLS" or "RADIUS/DTLS" is mentioned, it only applies to RADIUS/DTLS.

Changes to RADIUS This section discusses the needed changes to the RADIUS packet format (), port usage and shared secrets ().

Packet format The RADIUS packet format is unchanged from , and . Specifically, all of the following portions of RADIUS MUST be unchanged when using RadSec:

• Packet format
• Permitted codes
• Request Authenticator calculation
• Response Authenticator calculation
• Minimum packet length
• Maximum packet length
• Attribute format
• Vendor-Specific Attribute (VSA) format
• Permitted data types
• Calculation of dynamic attributes such as CHAP-Challenge, or Message-Authenticator
• Calculation of "encrypted" attributes such as Tunnel-Password.

The use of (D)TLS transport does not change the calculation of security-related fields (such as the Response-Authenticator) in RADIUS or RADIUS Dynamic Authorization . Calculation of attributes such as User-Password or Message-Authenticator also does not change. The changes to RADIUS implementations required to implement this specification are largely limited to the portions that send and receive packets on the network and the establishment of the (D)TLS connection. The requirement that RADIUS remain largely unchanged ensures the simplest possible implementation and widest interoperability of the specification. This includes the usage of the outdated security mechanisms in RADIUS that are based on shared secrets and MD5. This is not considered a security issue, since integrity and confidentiality are provided by the (D)TLS layer. See of this document or for more details. We note that for RADIUS/DTLS the DTLS encapsulation of RADIUS means that UDP datagrams include an additional overhead due to DTLS. This is discussed further in .

Default ports and shared secrets IANA has reserved ports for RADIUS/TLS and RADIUS/DTLS. Since authentication of peers, confidentiality, and integrity protection is achieved on the (D)TLS layer, the shared secret for the RADIUS packets is set to a static string, depending on the method. The calculation of security-related fields such as Response-Authenticator, Message-Authenticator or encrypted attributes MUST be performed using this shared secret.

Protocol	Port	Shared Secret
RADIUS/TLS	2083/tcp	"radsec"
RADIUS/DTLS	2083/udp	"radius/dtls"

RadSec does not use separate ports for authentication, accounting and dynamic authorization changes. The source port is arbitrary. For considerations regarding the multi-purpose use of one port for authentication and accounting see . RadSec endpoints MUST NOT use the old RADIUS/UDP or RADIUS/TCP ports for RADIUS/DTLS or RADIUS/TLS.

Detecting Live Servers RadSec implementations MUST utilize the existence of a TCP, TLS or DTLS connection where applicable in addition to the application-layer watchdog defined in when determining liveliness of each connection. As RADIUS is a "hop-by-hop" protocol, proxies hide information about the topology downstream to the client. While the client may be able to deduce the operational state of the next-hop (i.e. proxy), it is unable to determine the operational state of any hops beyond it. This is particularly problematic for topologies that aggregate multiple routes for differing realms behind a proxy where the absence of a reply could lead to a client to incorrectly deduce that the proxy is unavailable when the cause was an unresponsive downstream hop for a single realm. A similar effect may also be seen on home servers that uses different credential backends for each realm they service. To avoid these issues, RadSec clients MUST mark a connection 'DOWN' (as labelled by ) if one or more of the following conditions are met:

• The network stack indicates that the connection is no longer viable; such as the destination being no longer routable or the underlying TCP connection has been closed by the peer.
• The transport layer, D(TLS), provides no usable connection
• The application-layer watchdog algorithm has marked it 'DOWN'.

When a client opens multiple connections to a server, it is also possible that only one of the connections is unresponsive, e.g. because the server deleted the DTLS connection shared state or the connection was load balanced on the server side to a backend server that is now unresponsive. Therefore, the liveness check MUST be done on a per-connection basis, and a failure on one connection MUST NOT lead to all connections to this server being marked down. RadSec clients MUST implement the Status-Server extension as described in as the application level watchdog to detect the liveliness of the peer in the absence of responses. RadSec servers MUST be able to answer to Status-Server requests. Since RADIUS has a limitation of 256 simultaneous "in flight" packets due to the length of the ID field (), it is RECOMMENDED that RadSec clients reserve ID zero (0) on each connection for Status-Server packets. This value was picked arbitrarily, as there is no reason to choose any other value over another for this use. For RADIUS/TLS, the endpoints MAY send TCP keepalives as described in . For RADIUS/DTLS connections, the endpoints MAY send periodic keepalives as defined in . This is a way of proactively and rapidly triggering a connection DOWN notification from the network stack. These liveliness checks are essentially redundant in the presence of an application-layer watchdog, but may provide more rapid notifications of connectivity issues.

RadSec Packet and Connection Handling This section defines the behavior of RadSec endpoints for the handling of establishment of a (D)TLS connection and sending and receiving RADIUS packets. Server implementations MUST support both RADIUS/TLS and RADIUS/DTLS. Client implementations SHOULD implement both, but MUST implement at least one of RADIUS/TLS or RADIUS/DTLS.

(D)TLS requirements RadSec clients MUST establish a (D)TLS session immediately upon connecting to a new server. All data received over a TCP or UDP port assigned for RadSec is opaque for the RADIUS client or server application and must be handled by the TLS or DTLS implementation. Closing TLS connections and discarding invalid UDP datagrams are done by the (D)TLS implementation. RadSec has no notion of negotiating (D)TLS in an ongoing communication. As RADIUS has no provisions for capability signaling, there is also no way for a server to indicate to a client that it should transition to using TLS or DTLS. Servers and clients need to be preconfigured to use RADIUS/(D)TLS for a given endpoint. This action has to be taken by the administrators of the two systems. Implementations MUST follow the recommendations given in , especially in regards to recommended cipher suites and TLS session resumption. Additionally, the following requirements have to be met for the (D)TLS connection:

• Support for TLS 1.2 / DTLS 1.2 is REQUIRED, support for TLS 1.3 / DTLS 1.3 or higher is RECOMMENDED.
• Negotiation of a cipher suite providing for confidentiality as well as integrity protection is REQUIRED.
• The endpoints MUST NOT negotiate compression.
• The connection MUST be mutually authenticated (see )

The use of 0-RTT features of (D)TLS is NOT RECOMMENDED. RADIUS packets may contain confidential data that should be protected by forward secrecy, which 0-RTT cannot provide. If 0-RTT is used, implementations MUST also implement protection mechanisms against replay attacks. See for more detail.

Mutual authentication RadSec servers MUST authenticate clients, and RadSec clients MUST authenticate the server. RADIUS is designed to be used by mutually trusted systems. Allowing anonymous clients would ensure privacy for RadSec traffic, but would negate all other security aspects of the protocol, including security aspects of RADIUS itself, due to the fixed shared secret. RADIUS/(D)TLS allows for the following modes of mutual authentication, which will be further specified in this section:

• TLS-X.509-PKIX
• TLS-PSK

Independent of the chosen mode of authentication, the mutual authentication MUST be performed during the initial handshake. Alternative methods, such as post-handshake certificate-based client authentication (see ) with TLS 1.3 or renegotiation with TLS 1.2, MUST NOT be used to achieve mutual authentication.

Authentication using X.509 certificates with PKIX trust model (TLS-X.509-PKIX) All RadSec server implementations MUST implement this model. RadSec client implementations SHOULD implement this model, but MUST implement either this model or TLS-PSK. If implemented, the following rules apply:

• Implementations MUST allow the configuration of a trust base (i.e. a set of trusted Certificate Authorities (CAs)) for new TLS connections. This list SHOULD be application-specific and not use a global system trust store.
• Certificate validation MUST include the verification rules as per .
• Implementations SHOULD indicate their trust anchors when opening or accepting TLS connections. See and for TLS 1.2 and for TLS 1.3.
• When the configured trust base changes (e.g., removal of a CA from the set of trust anchors; issuance of a new CRL for a CA in the set of trust anchors), implementations SHOULD reassess the continued validity of the certificate path of all connected peers. This can either be done by caching the peer's certificate for the duration of the connection and re-evaluating the cached certificate or by renegotiating the (D)TLS connection, either directly or by opening a new (D)TLS connection and closing the old one.
• Implementations SHOULD NOT keep a connection open for longer than the validity span of the peer certificate. At the time the peer certificate expires, the connection SHOULD be closed and re-opened.

RadSec endpoints SHOULD NOT be pre-configured with a set of trusted CAs by the vendor or manufacturer that are enabled by default. Instead, the endpoints SHOULD start off with an empty CA set as the trust base. The addition of a CA SHOULD be done only when manually configured by the administrator. This does not preclude vendors or manufacturers including their set of trusted CAs in their products, but the enabling of those lists should be a conscious decision by an administrator. RadSec endpoints MUST follow when validating peer identities. Specific details are provided below:

• Certificates MAY use wildcards in the identifiers of DNS names and realm names, but only as the complete, left-most label.
• RadSec clients validate the server's identity to match their local configuration, accepting the identity on the first match:
  • If the expected RadSec server is associated with a specific NAI realm, e.g. by dynamic discovery or static configuration, that realm is matched against the presented identifiers of any subjectAltName entry of type otherName whose name form is NAIRealm as defined in .
  • If the expected RadSec server was configured as a hostname, or the hostname was yielded by a dynamic discovery procedure, that name is matched against the presented identifiers of any subjectAltName entry of type dNSName . Since a dynamic discovery might by itself not be secured, implementations MAY require the use of DNSSEC to ensure the authenticity of the DNS result before considering this identity as valid.
  • If the expected RadSec server was configured as an IP address, the configured IP address is matched against the presented identifier in any subjectAltName entry of type iPAddress .
  • The Common Name RDN MUST NOT be used to identify a server.
  • Clients MAY use other attributes of the certificate to validate the server's identity, but they MUST NOT accept any certificate without validation.
  • Clients which also act as servers (i.e. proxies) may be susceptible to security issues when a ClientHello is mirrored back to themselves. More details on this issue are discussed in .
• RadSec servers validate the certificate of the RadSec client against a local database of acceptable clients. The database may enumerate acceptable clients either by IP address or by a name component in the certificate.
  • For clients configured by DNS name, the configured name is matched against the presented identifiers of any subjectAltName entry of type dNSName .
  • For clients configured by their source IP address, the configured IP address is matched against the presented identifiers of any subjectAltName entry of type iPAddress .
  • Some servers MAY be configured to accept a client coming from a range or set of IP addresses. In this case, the server MUST verify that the client IP address of the current connection is a member of the range or set of IP addresses, and the server MUST match the client IP address of the current connection against the presented identifiers of any subjectAltName entry of type iPAddress .
  • Implementations MAY consider additional subjectAltName extensions to identify a client.
  • If configured by the administrator, the identity check MAY be omitted after a successful trust chain check, e.g. if the client used dynamic lookup there is no configured client identity to verify. The client's authorization MUST then be validated using a certificate policy OID unless both endpoints are part of a trusted network.
• Implementations MAY allow configuration of a set of additional properties of the certificate to check for a peer's authorization to communicate (e.g. a set of allowed values presented in subjectAltName entries of type uniformResourceIdentifier or a set of allowed X.509v3 Certificate Policies).

Authentication using TLS-PSK (TLS-PSK) RadSec server implementations MUST support the use of TLS-PSK. RadSec client implementations SHOULD support the use of TLS-PSK, but MUST implement either this model or TLS-X.509-PKIX. Further guidance on the usage of TLS-PSK in RadSec is given in .

Connecting Client Identity In RADIUS/UDP, clients are uniquely identified by their IP addresses, as the shared secret is associated with the origin IP address. With RadSec the shared secret has a fixed value and multiple distinct RadSec clients can connect from the same IP address. This requires changing the method of identifying individual clients from RADIUS/UDP. Depending on the trust model used, the RadSec client identity is determined as follows. With TLS-PSK, a client is uniquely identified by its TLS-PSK identifier. With TLS-X.509-PKIX, a client is uniquely identified by the tuple of the serial number of the presented client certificate and the issuer. In practice, identification of unique clients is not always necessary and could be based on the subject of the presented certificate or a subjectAltName entry. While this identification technique could match multiple distinct certificates and therefore distinct clients, it is often sufficient, e.g. for the purpose of applying policies. Note well: having identified a connecting entity does not mean the server necessarily wants to communicate with that client. For example, if the Issuer is not in a trusted set of Issuers, the server may decline to perform RADIUS transactions with this client. Additionally, a server MAY restrict individual or groups of clients to certain IP addresses or ranges. One example of this can be to restrict clients configured by DNS name to only the IP address(es) that this DNS name resolves to. A client connecting from outside the allowed range would be rejected, even if the mutual authentication otherwise would have been successful. To reduce server load and to prevent probing the validity of stolen credentials, the server SHOULD abort the (D)TLS handshake immediately with a TLS alert access_denied(49) after the client transmitted identifying information, i.e. the client certificate or the PSK identifier, and the server recognizes that the client connects from outside the allowed IP range. There are numerous trust models in PKIX environments, and it is beyond the scope of this document to define how a particular deployment determines whether a client is trustworthy. Implementations that want to support a wide variety of trust models should expose as many details of the presented certificate to the administrator as possible so that the trust model can be implemented by the administrator. As a suggestion, at least the following information from the TLS connection and the X.509 client certificate should be exposed:

• Originating IP address
• Certificate Fingerprint
• Issuer
• Subject
• all X.509v3 Extended Key Usage
• all X.509v3 Subject Alternative Name
• all X.509v3 Certificate Policy

In TLS-PSK operation at least the following information from the TLS connection should be exposed:

• Originating IP address
• TLS-PSK Identifier

TLS Session Resumption Session resumption lowers the time and effort required to start a (D)TLS connection and increases network responsiveness. This is especially helpful when using short idle timeouts. RadSec clients and servers SHOULD implement session resumption. Implementations supporting session resumption MUST cache data during the initial full handshake, sufficient to allow authorization decisions to be made during resumption. For RadSec servers, this should preferably be done using stateless session resumption as specified in , to reduce the resource usage for cached data. When establishing a (D)TLS connection with session resumption, both client and server MUST re-authorize the connection by using the original, cached data. In particular, this includes the X.509 certificate (when using a PKIX trust model) as well as any policies associated with that identity such as restrictions on source IP address. The re-authorization MUST give the same result as if a full handshake was performed at the time of resumption. If cached data cannot be retrieved securely, resumption MUST NOT be done, by either immediately closing the connection or reverting to a full handshake. If a connection that used session resumption is closed immediately after being established, the RadSec client MUST NOT re-attempt session resumption but perform a full TLS handshake instead.

RADIUS packets The RadSec specification does not change the client/server architecture of RADIUS. RadSec clients transmit the same packet types on the connection they initiated as a RADIUS/UDP client would, and RadSec servers transmit the same packet types on the connections the server has accepted as a RADIUS/UDP server would. As noted in , RadSec uses the same port for Authentication and Accounting packets. As non-exhaustive example, a RadSec client can transmit packets of type Access-Request, Accounting-Request, Status-Server, Disconnect-ACK over the same connection, and a RadSec server can transmit packets of type Access-Accept, Access-Reject, Access-Challenge, Accounting-Response, Disconnect-Request. However, special considerations apply for mixing Authentication and Accounting packets over the same connection. Traditional RADIUS/UDP uses different ports for Authentication and Accounting, where RadSec uses the same connection for all RADIUS packets. Due to the use of one single port for all packet types, clients might send packets to the server which it cannot process. Without a response from the server, the client has to wait for the requests to time out before reusing the request id, leading to resource exhaustion of the limited id space. A server MAY therefore respond with a Protocol-Error packet as defined in , to alleviate this situation and signal that it was unable to process a packet. The Error-Cause attribute of this packet SHOULD be set to the value 406 ("Unsupported Extension"), if the server does not support the packet type, or the value 502 ("Request Not Routable (Proxy)"), if the request cannot be routed. Future specifications may recommend other Error-Cause attribute values for specific scenarios. RadSec clients MUST accept Protocol-Error as a valid response and thus stop any retransmission of the original packet over the current connection. Further details of handling the Protocol-Error reply on the client side are outside of the scope of this document, see for a more detailed description on Protocol-Error.

Differences from RFC 6614 unwanted RADIUS packet handling The previous specification of RADIUS/TLS in recommends to send a reply that depends on the request type:

• For unwanted CoA-Requests or Disconnect-Requests, the servers should respond with a CoA-NAK or Disconnect-NAK, respectively.
• For unwanted Accounting-Requests, the servers should respond with an Accounting-Response containing an Error-Cause attribute with the value 406 ("Unsupported Extension").

also recommends that a RADIUS/TLS client observing this Accounting-Response should stop sending Accounting-Request packets to this server. This behavior, however, could lead to problems, especially in proxy fabrics, since the RADIUS client cannot determine whether the reply came from the correct server or a RADIUS proxy along the way. Compared to the recommended replies (CoA-NAK, Disconnect-NAK and Accounting-Response), the Protocol-Error packet is explicitly only applicable to one RADIUS hop and must not be forwarded, which gives the RADIUS client the opportunity to re-route the unwanted packet to a different RADIUS server. This also is backwards compatible with existing implementations, since RADIUS clients must ignore any incoming RADIUS packets with an unknown packet type. Therefore these recommended reply message types are now replaced with the Protocol-Error packet type.

Forwarding RADIUS packets between UDP and TCP based transports When a RADIUS proxy forwards packets, it is possible that the incoming and outgoing links have substantially different properties. This issue is most notable in UDP to TCP proxying, but there are still possible issues even when the same transport is used on both incoming and outgoing links. noted this issue many years ago: These differences are most notable in throughput, and in differing retransmission requirements.

Throughput Differences lead to Network Collapse An incoming link to the proxy may have substantially different throughput than the outgoing link. Perhaps the network characteristics on the two links are different, or perhaps the home server is slow. In both situations, the proxy may be left with a difficult choice about what to do with the incoming packets, if the rate of incoming packets exceeds throughput on the outgoing link. As RADIUS does not provide for connection-based congestion control, there is no way for the proxy to signal on the incoming link that the client should slow its rate of sending packets. As a result, the proxy must simply accept the packets, buffer them, and hope that they can be be sent outbound before the client gives up on the request.

Differing Retransmission Requirements Due to the lossy nature of UDP, RADIUS/UDP and RADIUS/DTLS transports are required to perform retransmissions as per . In contrast, RADIUS/TCP and RADIUS/TLS transports are reliable, and do not perform retransmissions. These requirements lead to an issue for proxies when they send packets across protocol boundaries with differing retransmission behaviors. When a proxy receives packets on an unreliable transport, and forwards them across a reliable transport, it receives retransmissions from the client, but MUST NOT forward those retransmissions across the reliable transport. The proxy MAY log information about these retransmissions, but it does not perform any other action. When a proxy receives RADIUS packets on a reliable transport, and forwards them across an unreliable transport, the proxy MUST perform retransmissions across the unreliable transport as per . That is, the proxy takes responsibility for the retransmissions. Implementations MUST take care to not completely decouple the two transports in this situation. See for details on retransmitting RADIUS packets over DTLS. That is, if an incoming connection on a reliable transport is closed, there may be pending retransmissions on an outgoing unreliable transport. Those retransmissions MUST be stopped, as there is nowhere to send the reply. Similarly, if the proxy sees that the client has given up on a request (such as by re-using an Identifier before the proxy has sent a response), the proxy MUST stop all retransmissions of the old request and discard it. The above requirements are a logical extension of the common practice where a client stops retransmission of a packet once it decides to "give up" on the packet and discard it. Whether this discard process is due to internal client decisions, or interaction with incoming connections is irrelevant. When the client cannot do anything with responses to a request, it MUST stop retransmitting that request.

Acct-Delay-Time and Event-Timestamp In order to avoid congestive collapse, it is RECOMMENDED that RadSec clients which originate Accounting-Request packets (i.e. not proxies) do not include Acct-Delay-Time () in those packets. Instead, those clients SHOULD include Event-Timestamp (), which is the time at which the original event occurred. The Event-Timestamp MUST NOT be updated on any retransmissions, as that would both negate the meaning of Event-Timestamp, and create the same problem as with Acct-Delay-Time. Not using Acct-Delay-Time allows for RADIUS packets to be retransmitted without change. In contrast, updating Acct-Delay-Time would require that the client create and send a new packet without signaling the server that the previous packet is no longer considered active. This process can occur repeatedly, which leads to multiple different packets containing effectively the same information (except for Acct-Delay-Time). This duplication contributes to congestive collapse of the network, if a RADIUS proxy performs retransmission to the next hop for each of those packets independently. Additionally, the different properties of the RADIUS/TLS transport as well as cross-protocol proxying change the assumption of a negligible transmission time of the RADIUS packet, on which the value of Acct-Delay-Time is based. While a single UDP packet may have a negligible transmission time, application data sent via TLS could arrive at the sender with a significant delay due to the underlying TCP retransmission mechanism. If the packet is proxied from RADIUS/TLS to RADIUS/DTLS or RADIUS/UDP, the proxy has to retransmit on its own without changing the value of Acct-Delay-Time, which again introduces non-negligible transmission delays. Using Event-Timestamp instead of Acct-Delay-Time also removes an ambiguity around retransmitted packets for RADIUS/TLS. Since there is no change to the packet contents when a retransmission timer expires, no new packet ID is allocated, and therefore no new packet is created. Where RadSec clients do include Acct-Delay-Time in RADIUS packets, the client SHOULD use timers to detect packet loss, as described in . RadSec clients SHOULD NOT update the Acct-Delay-Time, and therefore create a new RADIUS packet with the same information, until the timer has determined that the original packet has in fact been completely lost. This ensures that there is no congestive collapse, since a new packet is only created if following hops have also given up on retransmission, while keeping the functionality of Acct-Delay-Time to determine how long ago the event occurred. It only reduces the granularity of Acct-Delay-Time to the retransmission timeout, compared to the different approach of updating the Acct-Delay-Time on each retransmission.

Client Timers RadSec clients may need to reconnect to a server that rejected their connection attempt and retry RADIUS packets which did not get an answer. The following sections define the client behavior.

Reconnection attempts In contrast to RADIUS/UDP, RadSec establishes a (D)TLS connection before transmitting any RADIUS packets. Therefore, in addition to retransmission of RADIUS packets, RadSec clients also have to deal with connection retries. Except in cases where a connection attempt with session resumption was closed by the RadSec server, RadSec clients MUST NOT immediately reconnect to a server after a failed connection attempt. A connection attempt is treated as failed if it fails at any point until a (D)TLS connection is established successfully. Typical reconnections MUST have a lower bound for the time in between retries. The lower bound SHOULD be configurable, but MUST NOT be less than 0.5 seconds. In cases where the server closes the connection on an attempted TLS session resumption, the client MUST NOT use TLS session resumption for the following connection attempt. RadSec clients MUST implement an algorithm for handling the timing of such reconnection attempts that includes an exponential back-off. Using an algorithm similar to the retransmission algorithm defined in is RECOMMENDED. If a different algorithm is used, it SHOULD include a configurable lower and upper bound for the time between retries, a configurable timeout after which the client gives up reconnecting and a jitter. Where the connection to a RadSec server is established only when there is a RADIUS packet to be sent, adding a second RADIUS packet to be send SHOULD NOT trigger an immediate reconnection attempt. Instead, the algorithm SHOULD continue as it would have without the new packet, but the client MAY reset the timeout for giving up reconnecting. Where the connection to a RadSec server is configured to be static and always kept open, the reconnect algorithm SHOULD have an upper limit for the time between retries (e.g. 60 seconds) and not give up trying to reconnect.

RADIUS packet retransmission RadSec clients MUST implement retransmission timers for retransmitting RADIUS packets such as the ones defined in . Other algorithms than the one defined in are possible, but any timer implementations MUST have similar properties of including jitter, exponential backoff and a maximum retransmission count (MRC) or maximum retransmission duration (MRD). As TLS is a reliable transport, RADIUS/TLS clients can only retry a packet if a connection closes without that packet receiving a reply, therefore the timers MUST NOT result in retransmission of any packet. A retry is the re-sending of the same content in a newly constructed RADIUS packet, where a retransmission is the re-sending of the exact same packet over the same connection to deal with packet loss on transport. Instead, the timers, MRC or MRD specifically, can be used to determine that a packet will most likely not receive an answer ever, for example because a packet loss has occurred in a later RADIUS hop or the home server ignores the RADIUS packet. See for more discussion on retransmission behavior.

(D)TLS connection limits and timeout While RADIUS/UDP could be implemented mostly stateless (except for the requests in flight), both TCP/TLS as well as DTLS require state tracking of the underlying (D)TLS connection and are thus subject to potential resource exhaustion. This is aggravated by the fact that RADIUS client/servers are often statically configured and thus form long-running peer relationships with long-running connections. Implementations SHOULD have configurable limits on the number of open connections. When this maximum is reached and a new (D)TLS connection is needed, the server MUST either drop an old connection in order to open the new one or not create a new connection. The close notification of (D)TLS or underlying connections are not fully reliable, or connections might be unnecessarily kept alive by heartbeat or watchdog traffic, occupying resources. Therefore, both RadSec clients and servers MAY close connections after they have been idle for some time (no traffic except application layer watchdog). This idle timeout SHOULD be configurable within reasonable limits and it SHOULD be possible to disable idle timeouts completely. On the server side, this mostly helps avoid resource exhaustion. For clients, proactively closing connections can also help mitigate situations where watchdog mechanisms are unavailable or fail to detect non-functional connections. Some scenarios or RADIUS protocol extensions could also require that a connection be kept open at all times, so clients MAY immediately re-open the connection. These scenarios could be related to monitoring the infrastructure or to allow the server to proactively send packets to the clients without a preceding request. The value of the idle timeout to use depends on the exact deployment and is a trade-of between resource usage on clients/servers and the overhead of opening new connections. Very short timeouts that are at or below the timeouts used for application layer watchdogs, typically in the range of 30-60s can be considered unreasonable. In contrast, the upper limit is much more difficult to define but may be in the range of 10 to 15min, depending on the available resources, or never (disabling idle timeout) in scenarios where a permanently open connection is required.

Behavior on (D)TLS connection closure of incoming connections If an incoming (D)TLS connection or the underlying transport channel is closed or broken, then there is no way to send a RADIUS response packet to the client. The RadSec server behavior then depends on the types of packets being processed, and on the role of the server. A RadSec server acting as proxy MUST discard all requests associated with the closed connection. As no response can be sent over the now-closed (D)TLS connection, any further processing of requests is pointless. A discarded request may have a cached RADIUS response packet (), in which case the cached response also MUST be discarded. If there is no cached response packet, then the request might still be processed by the home server. The RADIUS proxy MUST discard any response to these requests and SHOULD stop processing the requests. A home server which receives Access-Request packets MUST behave as defined above for a proxy and discard those requests and stop processing them. Where a RADIUS packet is part of a multi-packet authentication session (e.g. EAP), the underlying authentication session could be continued, or the underlying authentication session data could be discarded. The server may be able to receive and process another packet for that session via a different incoming connection. It is difficult to make more recommendations for managing partially processed authentication sessions, as such recommendations depend strongly on the authentication method being used. As a result, further behavior is implementation defined and outside the scope of this specification. A home server which receives other kinds of packets (for example Accounting-Request, CoA-Request, Disconnect-Request) MAY finish processing outstanding requests, and then discard any response. This behavior ensures that the desired action is still taken, even if the home server cannot inform the client of the result of that action.

Malformed Packets and Unknown clients The RADIUS specifications say that an implementation should "silently discard" a packet in a number of circumstances. This action has no further consequences for UDP based transports, as the "next" packet is completely independent of the previous one. When TLS is used as transport, decoding the "next" packet on a connection depends on the proper decoding of the previous packet. As a result the behavior with respect to discarded packets has to change, since a malformed RADIUS packet could impact the decoding of succeeding packets. With DTLS, the "next" packet does not depend on proper decoding of the previous packet, since the RADIUS packets are sent in independent DTLS records. However, since both TLS and DTLS provide integrity protection and ensure that the packet was sent by the peer, a protocol violation at this stage implies that the peer is misbehaving. Implementations of this specification SHOULD treat the "silently discard" texts in the RADIUS specification referenced above as "silently discard and close the connection". That is, the implementation SHOULD send a TLS close notification and, in the case of RADIUS/TLS, the underlying TCP connection MUST be closed if any of the following circumstances are seen:

• Connection from an unknown client
• Packet where the RADIUS Length field is less than the minimum RADIUS packet length
• Packet where the RADIUS Length field is more than the maximum RADIUS packet length
• Packet where an Attribute Length field has the value of zero or one (0 or 1)
• Packet where the attributes do not exactly fill the packet
• Packet where the Request Authenticator fails validation (where validation is required)
• Packet where the Response Authenticator fails validation (where validation is required)
• Packet where the Message-Authenticator attribute fails validation (when it occurs in a packet)

After applying the above rules, there are still situations where the previous specifications allow a packet to be "silently discarded" upon receipt, but in which a connection MAY remain open:

• Packet with an invalid code field (see for details)
• Response packets that do not match any outstanding request
• A server lacking the resources to process a request

These requirements reduce the possibility for a misbehaving client or server to wreak havoc on the network.

RADIUS/TLS specific specifications This section discusses all specifications that are only relevant for RADIUS/TLS.

Sending and receiving RADIUS traffic The TLS layer of RADIUS/TLS provides a stream-based communication between the two peers instead of the traditional packet-based communication as with RADIUS/UDP. As a result, the way RADIUS packets are sent and received has to change. Instead of relying on packet borders of the underlying transport protocol to indicate the start of a new packet, the RADIUS/TLS endpoints have to keep track of the packet borders by examining the header of the received RADIUS packets. After the TLS connection is established, a RADIUS/TLS endpoint MUST NOT send any data except for RADIUS packets over the connection. Since the RADIUS packet header contains a Length field, the end of the RADIUS packet can be deduced. The next RADIUS packet MUST be sent directly after the RADIUS packet before, that is, the endpoints MUST NOT add padding before, between, or after RADIUS packets. When receiving RADIUS packets, a RADIUS/TLS endpoint MUST determine the borders of RADIUS packet based on the Length field in the RADIUS header. Note that, due to the stream architecture of TLS, it is possible that a RADIUS packet is first received only partially, and the remainder of the packet is contained in following fragments. Therefore, RADIUS/TLS endpoints MUST NOT assume that the packet length is invalid solely based on the currently available bytes in the stream. As an implementation note, it is RECOMMENDED that RADIUS/TLS implementations do not pass a single RADIUS packet to the TLS library in multiple fragments and instead assemble the RADIUS packet and pass it as one unit, in order to avoid unnecessary overhead when sending or receiving (especially if every new write generates a new TLS record) and wait times on the other endpoint.

Duplicates and Retransmissions As TCP is a reliable transport, RADIUS/TLS endpoints MUST NOT retransmit RADIUS packets over a given TCP connection. However, if the TLS connection or TCP connection is closed or broken, retries over new connections are permissible. RADIUS request packets that have not yet received a response MAY be transmitted by a RADIUS/TLS client over a new connection. As this procedure involves using a new connection, the ID of the packet MAY change. If the ID changes, any security attributes such as Message-Authenticator MUST be recalculated. Despite the above discussion, RADIUS/TLS servers SHOULD still perform duplicate detection on received packets, as described in . This detection can prevent duplicate processing of packets from non-conforming clients. RADIUS clients MUST NOT retry sending a packet by altering the protocol (i.e. switching from TLS to DTLS or vice versa) of the configured server on its own. This requirement does not, however, forbid the practice of putting servers with the same IP address and port but different protocols into a failover or load-balancing pool. In that situation, RADIUS requests MAY be retried with another server that is known to be part of the same pool.

TCP Applications Are Not UDP Applications Implementers should be aware that programming a robust TCP-based application can be very different from programming a robust UDP-based application. Additionally, differences in the transport like head-of-line blocking and the possibility of increased transmission times should be considered. When using RADIUS/UDP or RADIUS/DTLS, there is no ordering of packets. If a packet sent by a endpoint is lost, that loss has no effect on subsequent packets sent by that endpoint. Unlike UDP, TCP is subject to issues related to head-of-line blocking. This occurs when a TCP segment is lost and a subsequent TCP segment arrives out of order. While the RADIUS endpoints can process RADIUS packets out of order, the semantics of TCP makes this impossible. This limitation can lower the maximum packet processing rate of RADIUS/TLS. Additionally, due to the architecture of TCP as reliable stream transport, TCP retransmissions can occur significantly later, even multiple seconds, after the original data was passed to the network stack by the application. In contrast, RADIUS/UDP packets are usually received either quickly, or not at all, in which case the RADIUS/UDP stack triggers a retransmission of the packet on the application layer.

RADIUS/DTLS specific specifications This section discusses all specifications that are only relevant for RADIUS/DTLS.

RADIUS packet handling The DTLS encryption adds an additional overhead to each packet sent. RADIUS/DTLS implementations MUST support sending and receiving RADIUS packets of 4096 bytes in length, with a corresponding increase in the maximum size of the encapsulated DTLS packets. This larger packet size may cause the UDP packet to be larger than the Path MTU (PMTU), which causes the packet to be fragmented. Implementers and operators should be aware of the possibility of fragmented UDP packets. For details about issues with fragmentation see . RADIUS/DTLS endpoints MUST send exactly one RADIUS packet per DTLS record. This ensures that the RADIUS packets do not get fragmented at a point where a re-ordering of UDP packets would result in decoding failures. The DTLS specification mandates that a DTLS record must not span multiple UDP datagrams. We note that a single UDP datagram may, however, contain multiple DTLS records. RADIUS/DTLS endpoints MAY use this behavior to send multiple RADIUS packets in one UDP packet. For the receiving RADIUS/DTLS endpoint, the length checks defined in still apply. That is, a receiving RADIUS/DTLS endpoint MUST perform all the length checks, but MUST use the length of the decrypted payload of the DTLS record instead of the UDP packet length. Exactly one RADIUS packet is encapsulated in a DTLS record, and any data outside the range of the RADIUS length field within the decrypted payload of a single DTLS record MUST be treated as padding, as it would be with a RADIUS/UDP packet, and be ignored. For UDP datagrams containing multiple DTLS records, each DTLS record MUST be parsed individually. If a RADIUS packet should be re-transmitted, either as retransmission due to a missing response by the client or as retransmission of a cached response by the server, the RADIUS/DTLS endpoints MUST re-process the RADIUS packet through DTLS. That is, for the purpose of retransmissions, RADIUS/DTLS endpoints cache the RADIUS packet, as a RADIUS/UDP endpoint would, and not the DTLS record that contains the RADIUS packet.

Server behavior When a RADIUS/DTLS server receives packets on the configured RADIUS/DTLS port, all packets MUST be treated as being DTLS. RADIUS/UDP packets MUST NOT be accepted on this port. Some servers maintain a list of allowed clients per destination port. Others maintain a global list of clients that are permitted to send packets to any port. Where a client can send packets to multiple ports, the server MUST maintain a "DTLS Required" flag per client. This flag indicates whether or not the client is required to use DTLS. When set, the flag indicates that the only traffic accepted from the client is over the RADIUS/DTLS port. When packets are received from a client with the "DTLS Required" flag set on non-DTLS ports, the server MUST silently discard these packets, as there is no RADIUS/UDP shared secret available. This flag will often be set by an administrator. However, if the server receives DTLS traffic from a client, it SHOULD notify the administrator that DTLS is available for that client. It MAY mark the client as "DTLS Required". Allowing RADIUS/UDP and RADIUS/DTLS from the same client exposes the traffic to downbidding attacks and is NOT RECOMMENDED.

Client behavior When a RADIUS/DTLS client sends packet to the assigned RADIUS/DTLS port, all packets MUST be DTLS. RADIUS/UDP packets MUST NOT be sent to this port. RADIUS/DTLS clients SHOULD NOT probe servers to see if they support DTLS transport. Instead, clients SHOULD use DTLS as a transport layer only when administratively configured.

Session Management Where RADIUS/TLS can rely on the TCP state machine to perform session tracking, RADIUS/DTLS cannot. As a result, implementations of RADIUS/DTLS may need to perform session management of the DTLS session in the application layer. This subsection describes logically how this tracking is done. Implementations MAY choose to use the method described here, or another, equivalent method. When implementations do not use the 5-tuple described below, note that IP address based policies MUST still be applied for all incoming packets, similar to the mandated behavior for TLS Session Resumption in . We note that , already mandates a duplicate detection cache. The session tracking described below can be seen as an extension of that cache, where entries contain DTLS sessions instead of RADIUS/UDP packets.

Server Session Management A RADIUS/DTLS server using the 5-tuple method MUST track ongoing DTLS sessions for each client, based on the following 5-tuple:

• source IP address
• source port
• destination IP address
• destination port
• protocol (fixed to UDP)

Note that this 5-tuple is independent of IP address version (IPv4 or IPv6). Each 5-tuple points to a unique session entry, which usually contains the following information:

DTLS Session:

Any information required to maintain and manage the DTLS session.

DTLS Data:

An implementation-specific variable that may contain information about the active DTLS session. This variable may be empty or nonexistent.

This data will typically contain information such as idle timeouts, session lifetimes, and other implementation-specific data.

Session Opening and Closing Session tracking is subject to Denial-of-Service (DoS) attacks due to the ability of an attacker to forge UDP traffic. RADIUS/DTLS servers SHOULD use the stateless cookie tracking technique described in . DTLS sessions SHOULD NOT be tracked until a ClientHello packet has been received with an appropriate Cookie value. Server implementation SHOULD have a way of tracking DTLS sessions that are partially set up. Servers MUST limit both the number and impact on resources of partial sessions. Sessions (both 5-tuple and entry) MUST be deleted when the DTLS session is closed for any reason. When a session is deleted due to it failing security requirements, the DTLS session MUST be closed, any TLS session resumption parameters for that session MUST be discarded, and all tracking information MUST be deleted. Since UDP is stateless, the potential exists for the client to initiate a new DTLS session using a particular 5-tuple, before the server has closed the old session. For security reasons, the server MUST keep the old session active until either it has received secure notification from the client that the session is closed or the server decides to close the session based on idle timeouts. Taking any other action would permit unauthenticated clients to perform a DoS attack, by reusing a 5-tuple and thus causing the server to close an active (and authenticated) DTLS session. As a result, servers MUST ignore any attempts to reuse an existing 5-tuple from an active session. This requirement can likely be reached by simply processing the packet through the existing session, as with any other packet received via that 5-tuple. Non-compliant, or unexpected packets will be ignored by the DTLS layer.

Client Session Management RADIUS/DTLS clients SHOULD NOT send both RADIUS/UDP and RADIUS/DTLS packets to different servers from the same source socket. This practice causes increased complexity in the client application and increases the potential for security breaches due to implementation issues. RADIUS/DTLS clients MAY use PMTU discovery to determine the PMTU between the client and server, prior to sending any RADIUS traffic. While a RADIUS client has limited to no possibilities to reduce the size of an outgoing RADIUS packet without unwanted side effects, it gives the RADIUS client the possibility to determine whether or not the RADIUS packet can even be sent over the connection. IP fragmentation may not be functioning, so by determining the PMTU, the RADIUS client can preemptively select a different RADIUS server to send the RADIUS packet to. Further discussion of this problem is deemed outside of the scope of this document.

Security Considerations As this specification relies on the existing TLS and DTLS specifications, all security considerations for these protocols also apply to the (D)TLS portions of RadSec. For RADIUS however, many security considerations raised in the RADIUS documents are related to RADIUS encryption and authorization. Those issues are largely mitigated when (D)TLS is used as a transport method, since encryption and authorization is achieved on the (D)TLS layer. The issues that are not mitigated by this specification are related to the RADIUS packet format and handling, which is unchanged in this specification. A few remaining security considerations and notes to administrators deploying RadSec are listed below.

RADIUS Proxies RadSec provides authentication, integrity and confidentiality protection for RADIUS traffic between two RADIUS endpoints. In the presence of proxies, these intermediate proxies can still inspect the individual RADIUS packets, i.e., "end-to-end" encryption on the RADIUS layer is not provided. Where intermediate proxies are untrusted, it is desirable to use other RADIUS mechanisms to prevent RADIUS packet payload from inspection by such proxies. One common method to protect passwords is the use of the Extensible Authentication Protocol (EAP) and EAP methods that utilize TLS. Additionally, when RADIUS proxies are used, the RADIUS client has no way of ensuring that the complete path of the RADIUS packet is protected, since RADIUS routing is done hop-by-hop and any intermediate proxy may be configured, after receiving a RADIUS packet via RadSec from one endpoint, to forward this packet to a different endpoint using the RADIUS/UDP transport profile. There is no technical solution to this problem with the current specification. Where the confidentiality of the contents of the RADIUS packet across the whole path is required, organizational solutions need to be in place, that ensure that every intermediate RADIUS proxy is configured to forward the RADIUS packets using RadSec as transport. One possible way to reduce the attack surface is to reduce the number of proxies in the overall proxy chain. For this, dynamic discovery as defined in can be used.

Loopback-Attack on endpoints acting as Server and Client RadSec endpoints that are configured to act both as client and server, typically in a proxy configuration, may be vulnerable to attacks where an attacker mirrors back all traffic to the endpoint. Therefore, endpoints that are capable of acting as both client and server SHOULD implement mitigations to avoid accepting connections from itself. One example of a potentially vulnerable configuration is a setup where the RadSec server is accepting incoming connections from any address (or a wide address range). Since the server may not be able to verify the certificate subject or subject alternate names, the trust is based on the certificate issuer or certificate OID. However, in this case, the client certificate which the RadSec endpoint uses for outgoing connections on the client side might also satisfy the trust check of the server side. Other scenarios where the identification of an outgoing connection satisfies the trust check of an incoming one are possible, but are not enumerated here. Either through misconfiguration, erroneous or spoofed dynamic discovery, or an attacker rerouting TLS packets, a proxy might thus open a connection to itself, creating a loop. Such attacks have been described for TLS-PSK , dubbed a selfie-attack, but are much broader in the RadSec case. In particular, as described above, they also apply to certificate based authentication. Implementations SHOULD therefore detect connections from itself, and reject them. There is currently no detection method that works universally for all use-cases and TLS implementations. Some possible detection methods are listed below:

• Comparing client or server random used in the TLS handshake. While this is a very effective method, it requires access to values which are normally private to the TLS implementation.
• Sending a custom random number in an extension in the TLS client hello. Again, this is very effective, but requires extension of the TLS implementation.
• Comparing the incoming server certificate to all server certificates configured on the proxy. While in some scenarios this can be a valid detection method, using the same server certificate on multiple servers would keep these servers from connecting with each other, even when this connection is legitimate.

The application layer RADIUS protocol also offers some loop detection, e.g. using a Proxy-State attribute. However, these methods are not capable of reliably detecting and suppressing these attacks in every case and are outside the scope of this document.

Usage of null encryption cipher suites for debugging For debugging purposes, some TLS implementations offer cipher suites with NULL encryption, to allow inspection of the plaintext with packet sniffing tools. Since with RadSec the RADIUS shared secret is set to a static string ("radsec" for RADIUS/TLS, "radius/dtls" for RADIUS/DTLS), using a NULL encryption cipher suite will also result in complete disclosure of the whole RADIUS packet, including the encrypted RADIUS attributes, to any party eavesdropping on the conversation. Following the recommendations in , this specification forbids the usage of NULL encryption cipher suites for RadSec. For cases where administrators need access to the decrypted RadSec traffic, we suggest using different approaches, like exporting the key material from TLS libraries according to .

Possibility of Denial-of-Service attacks Both RADIUS/TLS and RADIUS/DTLS have a considerable higher amount of data that the server needs to store in comparison to RADIUS/UDP. Therefore, an attacker could try to exhaust server resources. With RADIUS/UDP, any bogus RADIUS packet would fail the cryptographic checks and the server would silently discard the bogus packet. For RadSec, the server needs to perform at least a partial TLS handshake to determine whether or not the client is authorized. Performing a (D)TLS handshake is more complex than the cryptographic check of a RADIUS packet. An attacker could try to trigger a high number of (D)TLS handshakes at the same time, resulting in a high server load and potentially a Denial-of-Service. To prevent this attack, a RadSec server SHOULD have configurable limits on new connection attempts. Both TLS and DTLS need to store connection information for each open (D)TLS connection. Especially with DTLS, a bogus or misbehaving client could open an excessive number of DTLS connections. This connection tracking could lead to a resource exhaustion on the server side, triggering a Denial-of-Service. Therefore, RadSec servers SHOULD have a configurable limit of the number of connections they can track. When the total number of connections tracked is going to exceed the configured limit, servers MAY free up resources by closing the connection that has been idle for the longest time. Doing so may free up idle resources that then allow the server to accept a new connection. RadSec servers MUST limit the number of partially open (D)TLS connections and SHOULD expose this limit as configurable option to the administrator. To prevent resource exhaustion by partially opening a large number of (D)TLS connections, RadSec servers SHOULD have a timeout on partially open (D)TLS connections. We recommend a limit of a few seconds, implementations SHOULD expose this timeout as configurable option to the administrator. If a (D)TLS connection is not established within this timeframe, it is likely that this is a bogus connection. In contrast, an established connection might not send packets for longer periods of time, but since the endpoints are mutually authenticated this does not pose a problem other than the problems mentioned before. A different means of prevention is IP filtering. If the IP range that the server expects clients to connect from is restricted, then the server can simply reject or drop all connection attempts from outside those ranges. If every RadSec client is configured with an IP range, then the server does not even have to perform a partial TLS handshake if the connection attempt comes from outside every allowed range, but can instead immediately drop the connection. To perform this lookup efficiently, RadSec servers SHOULD keep a list of the accumulated permitted IP address ranges, individually for each transport.

TLS Connection Lifetime and Key Rotation The RadSec TLS connections may have a long lifetime. Especially when dealing with high volume of RADIUS traffic, the encryption keys have to be rotated regularly, depending on both the amount of data which was transferred, and on the encryption method. See and for more information. Implementers SHOULD be aware of this issue and determine whether the underlying TLS library automatically rotates encryption keys or not. If the underlying TLS library does not perform the rotation automatically, RadSec implementations SHOULD perform this rotation manually, either by a key update of the existing TLS connection or by closing the TLS connection and opening a new one.

Connection Closing If malformed RADIUS packets are received or the packets fail the authenticator checks, this specification requires that the (D)TLS connection be closed. The reason is that the connection is expected to be used for transport of RADIUS packets only. Any non-RADIUS traffic on that connection means the other party is misbehaving and is potentially a security risk. Similarly, any RADIUS traffic failing authentication vector or Message-Authenticator validation means that two parties do not have a common shared secret. Since the shared secret is static, this again means the other party is misbehaving. We wish to avoid the situation where a third party can send well-formed RADIUS packets to a RADIUS proxy that cause a (D)TLS connection to close. Therefore, in other situations, the connection SHOULD remain open in the face of non-conforming packets. Any malformed RADIUS packets sent by a third party will go through the security checks of the RADIUS proxy upon reception and will not be forwarded. Well-formed RADIUS packets with portions that the proxy does not understand do not pose a security risk to the security properties of the RadSec connection and can be forwarded. This ensures forward compatibility with future RADIUS extensions.

Migrating from RADIUS/UDP to RadSec Since RADIUS/UDP security relies on the MD5 algorithm, which is considered insecure, using RADIUS/UDP over insecure networks is risky. We therefore recommend to migrate from RADIUS/UDP to RadSec. Within this migration process, however, there are a few items that need to be considered by administrators. Firstly, administrators may be tempted to simply migrate from RADIUS/UDP to RadSec with (D)TLS-PSK and reuse the RADIUS shared secret as (D)TLS-PSK. While this may seem like an easy way to upgrade RADIUS/UDP to RadSec, the cryptographic problems with the RADIUS/UDP shared secret render the shared secret potentially compromised. Using a potentially compromised shared secret as TLS-PSK compromises the whole TLS connection. Therefore, any shared secret used with RADIUS/UDP before MUST NOT be used with RadSec and (D)TLS-PSK. Implementers MUST NOT reuse the configuration option for the RADIUS/UDP shared secret for the (D)TLS-PSK to prevent accidental reuse. When upgrading from RADIUS/UDP to RadSec, there may be a period of time, where the connection between client and server is configured for both transport profiles. If the old RADIUS/UDP configuration is left configured, but not used in normal operation, e.g. due to a fail-over configuration that prefers RadSec, an attacker could disrupt the RadSec communication and force a downgrade to RADIUS/UDP. To prevent this it is RECOMMENDED that, when the migration to RadSec is completed, the RADIUS/UDP configuration is removed. RadSec clients MUST NOT fall back to RADIUS/UDP if the RadSec communication fails, unless explicitly configured this way. Special considerations apply for clients behind a NAT, where some clients use RADIUS/UDP and others use RadSec. A RADIUS server might not be able to detect if a RadSec client falls back to RADIUS/UDP, they will appear with the same source IP address to the server and use the same shared secret. It is therefore RECOMMENDED to not use RADIUS/UDP and RadSec clients behind a NAT at the same time.

Client Subsystems Many traditional clients treat RADIUS as subsystem-specific. That is, each subsystem on the client has its own RADIUS implementation and configuration. These independent implementations work for simple systems, but break down for RADIUS when multiple servers, fail-over and load-balancing are required. With (D)TLS enabled, these problems are expected to get worse. We therefore recommend in these situations to use a local proxy that arbitrates all RADIUS traffic between the client and all servers. This proxy will encapsulate all knowledge about servers, including security policies, fail-over and load-balancing. All client subsystems should communicate with this local proxy, ideally over a loopback address. The benefit of this configuration is that there is one place in the client that arbitrates all RADIUS traffic. Subsystems that do not implement RadSec can remain unaware of (D)TLS. (D)TLS connections opened by the proxy can remain open for a long period of time, even when client subsystems are restarted. The proxy can do RADIUS/UDP to some servers and RadSec to others. Delegation of responsibilities and separation of tasks are important security principles. By moving all RadSec knowledge to a (D)TLS-aware proxy, security analysis becomes simpler, and enforcement of correct security becomes easier.

Design Decisions Many of the design decisions of RADIUS/TLS and RADIUS/DTLS can be found in and . This section will discuss the rationale behind significant changes from the experimental specification.

Mandatory-to-implement transports With the merging of RADIUS/TLS and RADIUS/DTLS the question of mandatory-to-implement transports arose. In order to avoid incompatibilities, there were two possibilities: Either mandate one of the transports for all implementations or mandate the implementation of both transports for either the server or the client. As of the time writing, RADIUS/TLS is widely adapted for some use cases. However, TLS has some serious drawbacks when used for RADIUS transport. Especially the sequential nature of the connection and the connected issues like head-of-line blocking could create problems. Therefore, the decision was made that RADIUS servers must implement both transports. For RADIUS clients, that may run on more constrained hardware, implementers can choose to implement only the transport that is better suited for their needs.

Mandatory-to-implement trust profiles mandates the implementation of the trust profile "certificate with PKIX trust model" for both clients and servers. The experience of the deployment of RADIUS/TLS as specified in has shown that most actors still rely on RADIUS/UDP. Since dealing with certificates can create a lot of issues, both for implementers and administrators, for the re-specification we wanted to create an alternative to insecure RADIUS transports like RADIUS/UDP that can be deployed easily without much additional administrative overhead. As with the supported transports, the assumption is that RADIUS servers are generally believed to be less constrained than RADIUS clients. Since some client implementations already support using certificates for mutual authentication and there are several use cases, where pre-shared keys are not usable (e.g. a dynamic federation with changing members), the decision was made that, analog to the supported transports, RadSec servers must implement both certificates with PKIX trust model and TLS-PSK as means of mutual authentication. RadSec clients again can choose which method is better suited for them, but must, for compatibility reasons, implement at least one of the two.

Changes in application of TLS The original specification of RADIUS/TLS does not forbid the usage of compression in the TLS layer. As per , compression should not be used due to the possibility of compression-related attacks, unless the application protocol is proven to be not open to such attacks. Since some attributes of the RADIUS packets within the TLS tunnel contain values that an attacker could at least partially choose (i.e. username, MAC address or EAP message), there is a possibility for compression-related attacks, that could potentially reveal data in other RADIUS attributes through length of the TLS record. To circumvent this attack, this specification forbids the usage of TLS compression.

IANA Considerations Upon approval, IANA should update the Reference to radsec in the Service Name and Transport Protocol Port Number Registry:

• Service Name: radsec
• Port Number: 2083
• Transport Protocol: tcp/udp
• Description: Secure RADIUS Service
• Assignment notes: The TCP port 2083 was already previously assigned by IANA for "RadSec", an early implementation of RADIUS/TLS, prior to issuance of the experimental RFC 6614. [RFCXXXX] updates RFC 6614 (RADIUS/TLS) and RFC 7360 (RADIUS/DTLS).
• Reference: [RFCXXXX] (this document)

References Normative References This document specifies version 1.3 of the Transport Layer Security (TLS) protocol. TLS allows client/server applications to communicate over the Internet in a way that is designed to prevent eavesdropping, tampering, and message forgery. This document updates RFCs 5705 and 6066, and obsoletes RFCs 5077, 5246, and 6961. This document also specifies new requirements for TLS 1.2 implementations. This document specifies Version 1.2 of the Transport Layer Security (TLS) protocol. The TLS protocol provides communications security over the Internet. The protocol allows client/server applications to communicate in a way that is designed to prevent eavesdropping, tampering, or message forgery. [STANDARDS-TRACK] This document specifies the Transmission Control Protocol (TCP). TCP is an important transport-layer protocol in the Internet protocol stack, and it has continuously evolved over decades of use and growth of the Internet. Over this time, a number of changes have been made to TCP as it was specified in RFC 793, though these have only been documented in a piecemeal fashion. This document collects and brings those changes together with the protocol specification from RFC 793. This document obsoletes RFC 793, as well as RFCs 879, 2873, 6093, 6429, 6528, and 6691 that updated parts of RFC 793. It updates RFCs 1011 and 1122, and it should be considered as a replacement for the portions of those documents dealing with TCP requirements. It also updates RFC 5961 by adding a small clarification in reset handling while in the SYN-RECEIVED state. The TCP header control bits from RFC 793 have also been updated based on RFC 3168. This document specifies version 1.2 of the Datagram Transport Layer Security (DTLS) protocol. The DTLS protocol provides communications privacy for datagram protocols. The protocol allows client/server applications to communicate in a way that is designed to prevent eavesdropping, tampering, or message forgery. The DTLS protocol is based on the Transport Layer Security (TLS) protocol and provides equivalent security guarantees. Datagram semantics of the underlying transport are preserved by the DTLS protocol. This document updates DTLS 1.0 to work with TLS version 1.2. [STANDARDS-TRACK] This document specifies version 1.3 of the Datagram Transport Layer Security (DTLS) protocol. DTLS 1.3 allows client/server applications to communicate over the Internet in a way that is designed to prevent eavesdropping, tampering, and message forgery. The DTLS 1.3 protocol is based on the Transport Layer Security (TLS) 1.3 protocol and provides equivalent security guarantees with the exception of order protection / non-replayability. Datagram semantics of the underlying transport are preserved by the DTLS protocol. This document obsoletes RFC 6347. This document describes a protocol for carrying authentication, authorization, and configuration information between a Network Access Server which desires to authenticate its links and a shared Authentication Server. [STANDARDS-TRACK] This document describes a protocol for carrying accounting information between a Network Access Server and a shared Accounting Server. This memo provides information for the Internet community. This document describes a currently deployed extension to the Remote Authentication Dial In User Service (RADIUS) protocol, allowing dynamic changes to a user session, as implemented by network access server products. This includes support for disconnecting users and changing authorizations applicable to a user session. This memo provides information for the Internet community. In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. This document defines Remote Authentication Dial In User Service (RADIUS) support for the Extensible Authentication Protocol (EAP), an authentication framework which supports multiple authentication mechanisms. In the proposed scheme, the Network Access Server (NAS) forwards EAP packets to and from the RADIUS server, encapsulated within EAP-Message attributes. This has the advantage of allowing the NAS to support any EAP authentication method, without the need for method- specific code, which resides on the RADIUS server. While EAP was originally developed for use with PPP, it is now also in use with IEEE 802. This memo provides information for the Internet community. This document discusses transport issues that arise within protocols for Authentication, Authorization and Accounting (AAA). It also provides recommendations on the use of transport by AAA protocols. This includes usage of standards-track RFCs as well as experimental proposals. [STANDARDS-TRACK] This document describes a deployed extension to the Remote Authentication Dial In User Service (RADIUS) protocol, enabling clients to query the status of a RADIUS server. This extension utilizes the Status-Server (12) Code, which was reserved for experimental use in RFC 2865. This document is not an Internet Standards Track specification; it is published for informational purposes. This document formally deprecates Transport Layer Security (TLS) versions 1.0 (RFC 2246) and 1.1 (RFC 4346). Accordingly, those documents have been moved to Historic status. These versions lack support for current and recommended cryptographic algorithms and mechanisms, and various government and industry profiles of applications using TLS now mandate avoiding these old TLS versions. TLS version 1.2 became the recommended version for IETF protocols in 2008 (subsequently being obsoleted by TLS version 1.3 in 2018), providing sufficient time to transition away from older versions. Removing support for older versions from implementations reduces the attack surface, reduces opportunity for misconfiguration, and streamlines library and product maintenance. This document also deprecates Datagram TLS (DTLS) version 1.0 (RFC 4347) but not DTLS version 1.2, and there is no DTLS version 1.1. This document updates many RFCs that normatively refer to TLS version 1.0 or TLS version 1.1, as described herein. This document also updates the best practices for TLS usage in RFC 7525; hence, it is part of BCP 195. Transport Layer Security (TLS) and Datagram Transport Layer Security (DTLS) are used to protect data exchanged over a wide range of application protocols and can also form the basis for secure transport protocols. Over the years, the industry has witnessed several serious attacks on TLS and DTLS, including attacks on the most commonly used cipher suites and their modes of operation. This document provides the latest recommendations for ensuring the security of deployed services that use TLS and DTLS. These recommendations are applicable to the majority of use cases. RFC 7525, an earlier version of the TLS recommendations, was published when the industry was transitioning to TLS 1.2. Years later, this transition is largely complete, and TLS 1.3 is widely available. This document updates the guidance given the new environment and obsoletes RFC 7525. In addition, this document updates RFCs 5288 and 6066 in view of recent attacks. The specification for enhanced mail system status codes, RFC 3463, establishes a new code model and lists a collection of status codes. While it anticipated that more codes would be added over time, it did not provide an explicit mechanism for registering and tracking those codes. This document specifies an IANA registry for mail system enhanced status codes, and initializes that registry with the codes so far established in published standards-track documents, as well as other codes that have become established in the industry. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. This memo profiles the X.509 v3 certificate and X.509 v2 certificate revocation list (CRL) for use in the Internet. An overview of this approach and model is provided as an introduction. The X.509 v3 certificate format is described in detail, with additional information regarding the format and semantics of Internet name forms. Standard certificate extensions are described and two Internet-specific extensions are defined. A set of required certificate extensions is specified. The X.509 v2 CRL format is described in detail along with standard and Internet-specific extensions. An algorithm for X.509 certification path validation is described. An ASN.1 module and examples are provided in the appendices. [STANDARDS-TRACK] This document provides specifications for existing TLS extensions. It is a companion document for RFC 5246, "The Transport Layer Security (TLS) Protocol Version 1.2". The extensions specified are server_name, max_fragment_length, client_certificate_url, trusted_ca_keys, truncated_hmac, and status_request. [STANDARDS-TRACK] Many application technologies enable secure communication between two entities by means of Transport Layer Security (TLS) with Internet Public Key Infrastructure using X.509 (PKIX) certificates. This document specifies procedures for representing and verifying the identity of application services in such interactions. This document obsoletes RFC 6125. This document specifies a means to find authoritative RADIUS servers for a given realm. It is used in conjunction with either RADIUS over Transport Layer Security (RADIUS/TLS) or RADIUS over Datagram Transport Layer Security (RADIUS/DTLS). The Domain Name System Security Extensions (DNSSEC) add data origin authentication and data integrity to the Domain Name System. This document introduces these extensions and describes their capabilities and limitations. This document also discusses the services that the DNS security extensions do and do not provide. Last, this document describes the interrelationships between the documents that collectively describe DNSSEC. [STANDARDS-TRACK] This document describes a mechanism that enables the Transport Layer Security (TLS) server to resume sessions and avoid keeping per-client session state. The TLS server encapsulates the session state into a ticket and forwards it to the client. The client can subsequently resume a session using the obtained ticket. This document obsoletes RFC 4507. [STANDARDS-TRACK] The RADIUS-over-TLS experiment described in RFC 6614 has opened RADIUS to new use cases where the 4096-octet maximum size limit of a RADIUS packet proves problematic. This specification extends the RADIUS-over-TCP experiment (RFC 6613) to permit larger RADIUS packets. This specification compliments other ongoing work to permit fragmentation of RADIUS authorization information. This document registers a new RADIUS code, an action that required IESG approval. This document describes common issues seen in Remote Authentication Dial In User Service (RADIUS) implementations and suggests some fixes. Where applicable, ambiguities and errors in previous RADIUS specifications are clarified. [STANDARDS-TRACK] This document describes the Heartbeat Extension for the Transport Layer Security (TLS) and Datagram Transport Layer Security (DTLS) protocols. The Heartbeat Extension provides a new protocol for TLS/DTLS allowing the usage of keep-alive functionality without performing a renegotiation and a basis for path MTU (PMTU) discovery for DTLS. [STANDARDS-TRACK] Informative References This document describes the MD5 message-digest algorithm. The algorithm takes as input a message of arbitrary length and produces as output a 128-bit "fingerprint" or "message digest" of the input. This memo provides information for the Internet community. It does not specify an Internet standard. This document defines Application-Layer Protocol Negotiation (ALPN) extensions for use with RADIUS/TLS and RADIUS/DTLS. These extensions permit the negotiation of an application protocol variant of RADIUS called "RADIUS/1.1". No changes are made to RADIUS/UDP or RADIUS/TCP. The extensions allow the negotiation of a transport profile where the RADIUS shared secret is no longer used, and all MD5-based packet authentication and attribute obfuscation methods are removed. This document updates RFCs 2865, 2866, 5176, 6613, 6614, and 7360. Independent This document specifies version 1.3 of the Transport Layer Security (TLS) protocol. TLS allows client/server applications to communicate over the Internet in a way that is designed to prevent eavesdropping, tampering, and message forgery. This document updates RFCs 5705, 6066, 7627, and 8422 and obsoletes RFCs 5077, 5246, 6961, 8422, and 8446. This document also specifies new requirements for TLS 1.2 implementations. This document provides implementation and operational considerations for using TLS Pre-Shared Keys (TLS-PSKs) with RADIUS/TLS (RFC 6614) and RADIUS/DTLS (RFC 7360). The purpose of the document is to help smooth the operational transition from the use of RADIUS/UDP to RADIUS/TLS. InkBridge Networks We extend and standardise the Protocol-Error packet Code, first defined in RFC 7930 for the Remote Authentication Dial In User Service (RADIUS) protocol. This document describes additional attributes for carrying authentication, authorization and accounting information between a Network Access Server (NAS) and a shared Accounting Server using the Remote Authentication Dial In User Service (RADIUS) protocol described in RFC 2865 and RFC 2866. This memo provides information for the Internet community. This document describes IP fragmentation and explains how it introduces fragility to Internet communication. This document also proposes alternatives to IP fragmentation and provides recommendations for developers and network operators. This document provides usage guidance for external Pre-Shared Keys (PSKs) in Transport Layer Security (TLS) 1.3 as defined in RFC 8446. It lists TLS security properties provided by PSKs under certain assumptions, then it demonstrates how violations of these assumptions lead to attacks. Advice for applications to help meet these assumptions is provided. This document also discusses PSK use cases and provisioning processes. Finally, it lists the privacy and security properties that are not provided by TLS 1.3 when external PSKs are used. Mozilla Zscaler University of Applied Sciences Bonn-Rhein-Sieg A format that supports logging information about the secrets used in a TLS connection is described. Recording secrets to a file in SSLKEYLOGFILE format allows diagnostic and logging tools that use this file to decrypt messages exchanged by TLS endpoints. This format is intended for use in systems where TLS only protects test data. IBM Research Europe - Zurich Mozilla Cloudflare An Authenticated Encryption with Associated Data (AEAD) algorithm provides confidentiality and integrity. Excessive use of the same key can give an attacker advantages in breaking these properties. This document provides simple guidance for users of common AEAD functions about how to limit the use of keys in order to bound the advantage given to an attacker. It considers limits in both single- and multi-key settings. The RADIUS protocol defined in RFC 2865 has limited support for authentication and encryption of RADIUS packets. The protocol transports data in the clear, although some parts of the packets can have obfuscated content. Packets may be replayed verbatim by an attacker, and client-server authentication is based on fixed shared secrets. This document specifies how the Datagram Transport Layer Security (DTLS) protocol may be used as a fix for these problems. It also describes how implementations of this proposal can coexist with current RADIUS systems. This document specifies a transport profile for RADIUS using Transport Layer Security (TLS) over TCP as the transport protocol. This enables dynamic trust relationships between RADIUS servers. [STANDARDS-TRACK] The Remote Authentication Dial-In User Server (RADIUS) protocol has, until now, required the User Datagram Protocol (UDP) as the underlying transport layer. This document defines RADIUS over the Transmission Control Protocol (RADIUS/TCP), in order to address handling issues related to RADIUS over Transport Layer Security (RADIUS/TLS). It permits TCP to be used as a transport protocol for RADIUS only when a transport layer such as TLS or IPsec provides confidentiality and security. This document defines an Experimental Protocol for the Internet community.

Changes from RFC6614 (RADIUS/TLS) and RFC7360 (RADIUS/DTLS) The following list contains the most important changes from the previous specifications in (RADIUS/TCP), (RADIUS/TLS) and (RADIUS/DTLS).

• The protocols RADIUS/TLS and RADIUS/DTLS are now collectively referred to as RadSec.
• referenced for TCP-related specification, RFC6613 on the other hand had some specification for RADIUS/TLS. These specifications have been merged into this document, and therefore removes as normative reference.
• RFC6614 marked TLSv1.1 or later as mandatory, this specification requires TLSv1.2 as minimum and recommends usage of TLSv1.3.
• RFC6614 allowed use of TLS compression, this document forbids it.
• RFC6614 only requires support for the trust model "certificates with PKIX" (). This document changes this. For servers, TLS-X.509-PKIX (, equivalent to "certificates with PKIX" in RFC6614) and TLS-PSK () is now mandated and clients must implement at least one of the two.
• The recommendation for TLS-X509-FINGERPRINT () is removed since the model has not been implemented by any known implementation of the experimental RADIUS/(D)TLS specifications.
• The mandatory-to-implement cipher suites are not referenced directly, this is replaced by a pointer to the TLS BCP.
• The specification regarding steps for certificate verification has been updated.
• mandated the use of Status-Server as watchdog algorithm, only recommended it. This specification mandates the use of Status-Server for both RADIUS/TLS and RADIUS/DTLS.
• only included limited text around retransmissions, this document now gives more guidance on how to handle retransmissions, especially across different transports.
• The rules for verifying the peer certificate have been updated to follow guidance provided in . Using the Common Name RDN for validation of server certificates is now forbidden.
• The response to unwanted packets has changed. Endpoints should now reply with a Protocol-Error packet, which is connection-specific and should not be proxied.

The rationales behind some of these changes are outlined in .

Acknowledgments Thanks to the original authors of RFC 6613, RFC 6614 and RFC 7360: Alan DeKok, Stefan Winter, Mike McCauley, Stig Venaas and Klaas Vierenga. Thanks to Arran Curdbard-Bell for text around keepalives and the Status-Server watchdog algorithm. Thanks to Alan DeKok for his constant review of this document over its whole process and his many text contributions, like text around forwarding issues between TCP and UDP based transports.