RADIUS and TLS-PSK

Introduction The previous specifications "Transport Layer Security (TLS) Encryption for RADIUS" and "Datagram Transport Layer Security (DTLS) as a Transport Layer for RADIUS" defined how (D)TLS can be used as a transport protocol for RADIUS. However, those documents do not provide guidance for using TLS-PSK with RADIUS. This document provides that missing guidance, and gives implementation and operational considerations. Unless it is explicitly called out that a recommendation applies to TLS alone or to DTLS alone, each recommendation applies to both TLS and DTLS. This document uses "shared secret" to mean "RADIUS shared secret", and Pre-Shared Key (PSK) to mean secrets which are used with TLS-PSK.

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.

History TLS deployments usually rely on certificates in most common uses. However, we recognize that it may be difficult to fully upgrade client implementations to allow for certificates to be used with RADIUS/TLS and RADIUS/DTLS. These upgrades involve not only implementing TLS, but can also require significant changes to administration interfaces and application programming interfaces (APIs) in order to fully support certificates. For example, unlike shared secrets, certificates expire. This expiration means that a working system using TLS can suddenly stop working. Managing this expiration can require additional notification APIs on RADIUS clients and servers which were previously not required when shared secrets were used. Certificates also require the use of certification authorities (CAs), and chains of certificates. RADIUS implementations using TLS therefore have to track not just a small shared secret, but also potentially many large certificates. The use of TLS-PSK can therefore provide a simpler upgrade path for implementations to transition from RADIUS shared secrets to TLS.

General Discussion of PSKs and PSK Identities Before we define any RADIUS-specific use of PSKs, we must first review the current standards for PSKs, and give general advice on PSKs and PSK identities. The requirements in this section apply to both client and server implementations which use TLS-PSK. Client-specific and server-specific issues are discussed in more detail later in this document.

Requirements on PSKs Reuse of a PSK in multiple versions of TLS (e.g. TLS 1.2 and TLS 1.3) is considered unsafe (). Where TLS 1.3 binds the PSK to a particular key derivation function, TLS 1.2 does not. This binding means that it is possible to use the same PSK in different hashes, leading to the potential for attacking the PSK by comparing the hash outputs. While there are no known insecurities, these uses are not known to be secure, and should therefore be avoided. adds a key derivation function to the import interface of (D)TLS 1.3, which binds the externally provided PSK to the protocol version. In particular, that document:

... describes a mechanism for importing PSKs derived from external PSKs by including the target KDF, (D)TLS protocol version, and an optional context string to ensure uniqueness. This process yields a set of candidate PSKs, each of which are bound to a target KDF and protocol, that are separate from those used in (D)TLS 1.2 and prior versions. This expands what would normally have been a single PSK and identity into a set of PSKs and identities.

An implementation MUST NOT use the same PSK for TLS 1.3 and for earlier versions of TLS. This requirement prevents reuse of a PSK with multiple TLS versions, which prevents the attacks discussed in . The exact manner in which this requirement is enforced is implementation-specific. One possibility is to have two different PSKs. Another possibility is to forbid the use of TLS 1.3, or to forbid the use of TLS versions less than TLS 1.3. It is RECOMMENDED that systems follow the directions of for the use of external PSKs in TLS. That document provides extremely useful guidance on generating and using PSKs. Implementations MUST support PSKs of at least 32 octets in length, and SHOULD support PSKs of 64 octets or more. As the PSKs are generally hashed before being used in TLS, the useful entropy of a PSK is limited by the size of the hash output. This output may be 256, 384, or 512 bits in length. Never the less, it is good practice for implementations to allow entry of PSKs of more than 64 octets, as the PSK may be in a form other than bare binary data. Implementations which limit the PSK to a maximum of 64 octets are likely to use PSKs which have much less than 512 bits of entropy. That is, a PSK with high entropy may be expanded via some construct (e.g. base32 as in the example below) in order to make it easier for people to interact with. Where 512 bits of entropy are input to an encoding construct, the output may be larger than 64 octets. Implementations MUST require that PSKs be at least 16 octets in length, which SHOULD be derived from a source with at least 128 bits of entropy. That is, short PSKs MUST NOT be permitted to be used, and PSKs MUST be uniformly random. The strength of the PSK is not determined by the length of the PSK, but instead by the number of bits of entropy which it contains. People are not good at creating data with high entropy, so a source of cryptographically secure random numbers MUST be used. Administrators SHOULD use PSKs of at least 24 octets, generated using a source of cryptographically secure random numbers. Implementers needing a secure random number generator should see for for further guidance. PSKs are not passwords, and administrators should not try to manually create PSKs. Passwords are generally intended to be remembered and entered by people on a regular basis. In contrast, PSKs are intended to be entered once, and then automatically saved in a system configuration. As such, due to the limited entropy of passwords, they are not acceptable for use with TLS-PSK, and would only be acceptable for use with a password-authenticated key exchange (PAKE) TLS method . We also incorporate by reference the requirements of when using PSKs. In order to guide Implementers, we give an example script below which generates random PSKs. While the script is not portable to all possible systems, the intent here is to document a concise and simple method for creating PSKs which are both secure, and humanly manageable. This script reads 128 bits (16 octets) of random data from a secure source, encodes it in Base32, and then formats it to be more humanly manageable. The generated keys are of the form "yttb-4gv2-ynfk-jbjh-2dja-cj7e-am". This form of PSK will be accepted by any implementation which supports at least 32 octets for PSKs. Larger PSKs can be generated by passing larger values to the "urandom()" function. The above derivation assumes that the random source returns one bit of entropy for every bit of randomness which is returned. Sources failing that assumption are NOT RECOMMENDED.

Interaction between PSKs and RADIUS Shared Secrets Any shared secret used for RADIUS/UDP or RADIUS/TLS MUST NOT be used for TLS-PSK. It is RECOMMENDED that RADIUS clients and servers track all used shared secrets and PSKs, and then verify that the following requirements all hold true:

no shared secret is used for more than one RADIUS client
no PSK is used for more than one RADIUS client
no shared secret is used as a PSK

Note that the shared secret of "radsec" given in can be used across multiple clients, as that value is mandated by the specification. The intention here is to recommend best practices for administrators who enter site-local shared secrets. There may be use-cases for using one shared secret across multiple RADIUS clients. There may similarly be use-cases for sharing a PSK across multiple RADIUS clients. Details of the possible attacks on reused PSKs are given in . There are few, if any, use-cases for using a PSK as a shared secret, or vice-versa. Implementations SHOULD NOT provide user interfaces which allow both PSKs and shared secrets to be entered at the same time. There is too much of a temptation for administrators to enter the same value in both fields, which would violate the limitations given above. Implementations MUST NOT use a "shared secret" field as a way for administrators to enter PSKs. The PSK entry fields MUST be labeled as being related to PSKs, and not to shared secrets.

PSK Identities requires that PSK identities be encoded in UTF-8 format. However, reuses PSKs and the PSK identity for resumption, and defines the ticket as:

the value of the ticket to be used as the PSK identity. The ticket itself is an opaque label. It MAY be either a database lookup key or a self-encrypted and self-authenticated value.

These definitions appear to be in conflict. This conflict is addressed in , which discusses requirements for encoding and comparison of PSK identities. It is RECOMMENDED that systems follow the directions of when using PSK Identities for RADIUS/TLS. In general, implements should allow for administratively provisioned PSK identities to follow and be UTF-8, while PSK identities provisioned as part of resumption are automatically provisioned, and therefore follow . Note that the PSK identity is sent in the clear, and is therefore visible to attackers. Where privacy is desired, the PSK identity could be either an opaque token generated cryptographically, or perhaps in the form of a Network Access Identifier (NAI) , where the "user" portion is an opaque token. For example, an NAI could be "68092112@example.com". If the attacker already knows that the client is associated with "example.com", then using that domain name in the PSK identity offers no additional information. In contrast, the "user" portion needs to be both unique to the client and private, so using an opaque token there is a more secure approach. Implementations MUST support PSK Identities of 128 octets, and SHOULD support longer PSK identities. We note that while TLS provides for PSK identities of up to 2^16-1 octets in length, there are few practical uses for extremely long PSK identities. It is up to administrators and implementations as to how they differentiate administratively provisioned PSK identities from automatically provisioned identities used in TLS 1.3 session tickets. One approach could be to have administratively provisioned identities contain an NAI such as described above, while session tickets contain large blobs of opaque, encrypted, and authenticated text. It should then be relatively straightforward to differentiate the two types of identities. One is UTF-8, the other is not. One is not authenticated, the other is authenticated. Servers MUST assign and/or track session resumption identities in a way which facilities the ability to distinguish those identities from pre-configured ones ,and which enables them to both find and validate the session resumption ticket. A sample validation flow for TLS-PSK identities could be as follows:

PSK identities provided via an administration interface are enforced to be only UTF-8 on both client and server The client treats session tickets received from the server as opaque blobs When the server issues session tickets for resumption, the server ensures that they are not valid UTF-8 One way to do this is to use stateless resumption with a forced non-UTF-8 key_name per , such as by setting one octet to 0x00. When receiving TLS, the server receives Client-Hello containing a PSK, and checks if the identity is valid UTF-8. > If yes, it searches for a pre-configured client which matches that identity >> If the identity is found, authenticates the client via PSK >> else the identity is invalid, and the server closes the connection. > If the identity is not UTF-8, try resumption, which can be handled by a TLS library >> If the TLS library verifies the session ticket, resumption has happened, and the connection is established. >> else the server ignores the session ticker, and performs normal TLS handshake with a certificate.

This validation flow is only suggested. Other validation methods are possible.

Security of PSK Identities We note that the PSK identity is a field created by the connecting client. Since the client is untrusted until both the identity and PSK have been verified, both of those fields MUST be treated as untrusted. That is, a well-formed PSK Identity is likely to be in UTF-8 format, due to the requirements of . However, implementations MUST support managing PSK identities as a set of undistinguished octets. It is not safe to use a raw PSK Identity to look up a corresponding PSK. The PSK may come from an untrusted source, and may contain invalid or malicious data. For example, the identity may have incorrect UTF-8 format; or it may contain data which forms an injection attack for SQL, LDAP, REST or shell meta characters; or it may contain embedded NUL octets which are incompatible with APIs which expect NUL terminated strings. The identity may also be up to 65535 octets long. As such, implementations MUST validate the identity prior to it being used as a lookup key. When the identity is passed to an external API (e.g. database lookup), implementations MUST either escape any characters in the identity which are invalid for that API, or else reject the identity entirely. The exact form of any escaping depends on the API, and we cannot document all possible methods here. However, a few basic validation rules are suggested, as outlined below. Any identity which is rejected by these validation rules SHOULD cause the server to close the TLS connection. The suggested validation rules for identities used outside of resumption are as follows:

Identities longer than a fixed maximum SHOULD be rejected. The limit is implementation dependent, but SHOULD NOT be less than 128, and SHOULD NOT be more than 1024.
Identities SHOULD be in UTF-8 format. Identities with embedded control characters, NUL octets, etc. SHOULD NOT be used.
Where the NAI format is expected, identities which are not in NAI format SHOULD be rejected

It is RECOMMENDED that implementations extend these rules with any additional validation which are found to be useful. For example, implementations and/or deployments could both generate PSK identities in a particular format for passing to client systems, and then also verify that any received identity matches that format. For example, a site could generate PSK identities which are composed of characters in the local language. The site could then reject identities which contain characters from other languages, even if those characters are valid UTF-8.

PSK and PSK Identity Sharing While administrators may desire to share PSKs and/or PSK identities across multiple systems, such usage is NOT RECOMMENDED. Details of the possible attacks on reused PSKs are given in . Implementations MUST be able to configure a unique PSK and PSK identity for each possible client-server relationship. This configuration allows administrators desiring security to use unique PSKs for each such relationship. This configuration also allows administrators to re-use PSKs and PSK Identities where local policies permit. Implementations SHOULD warn administrators if the same PSK identity and/or PSK is used for multiple client-server relationships.

Guidance for RADIUS Clients Client implementations MUST allow the use of a pre-shared key (TLS-PSK) for RADIUS/TLS. The client implementation can then expose a user interface flag which is "TLS yes / no", and then also fields which ask for the PSK identity and PSK itself. For TLS 1.3, Implementations MUST support "psk_dhe_ke" Pre-Shared Key Exchange Mode in TLS 1.3 as discussed in and in . Implementations MUST implement the recommended cipher suites in for TLS 1.2, and in for TLS 1.3. In order to future-proof these recommendations, we give the following recommendations:

Implementations SHOULD use the "Recommended" cipher suites listed in the IANA "TLS Cipher Suites" registry,
- for TLS 1.3, the use "psk_dhe_ke" PSK key exchange mode,
- for TLS 1.2 and earlier, use cipher suites which require ephemeral keying.

If a client initiated a connection using a PSK with TLS 1.3 by including the pre-shared key extension, it MUST close the connection if the server did not also select the pre-shared key to continue the handshake.

PSK Identities is silent on the subject of PSK identities, which is an issue that we correct here. Guidance is required on the use of PSK identities, as the need to manage identities associated with PSK is a new requirement for NAS management interfaces, and is a new requirement for RADIUS servers. RADIUS systems implementing TLS-PSK MUST support identities as per , and MUST enable configuring TLS-PSK identities in management interfaces as per . The historic methods of signing RADIUS packets have not yet been cracked, but they are believed to be much less secure than modern TLS. Therefore, when a RADIUS shared secret is used to sign RADIUS/UDP or RADIUS/TCP packets, that shared secret MUST NOT be used with TLS-PSK. If the secrets were to be reused, then an attack on historic RADIUS cryptography could be trivially leveraged to decrypt TLS-PSK sessions. Therefore in order to prevent confusion between shared secrets and TLS-PSKs, management interfaces and APIs need to label PSK fields as "PSK" or "TLS-PSK", rather than as "shared secret". With TLS-PSK, RADIUS/TLS clients MUST permit the configuration of a RADIUS server IP address or host name, because dynamic server lookups can only be used if servers use certificates.

Guidance for RADIUS Servers In order to support clients with TLS-PSK, server implementations MUST allow the use of a pre-shared key (TLS-PSK) for RADIUS/TLS. For TLS 1.3, Implementations MUST support "psk_dhe_ke" Pre-Shared Key Exchange Mode in TLS 1.3 as discussed in and in . Implementations MUST implement the recommended cipher suites in for TLS 1.2, and in for TLS 1.3. In order to future-proof these recommendations, we give the following recommendations:

Implementations SHOULD use the "Recommended" cipher suites listed in the IANA "TLS Cipher Suites" registry,
- for TLS 1.3, the use "psk_dhe_ke" PSK key exchange mode,
- for TLS 1.2 and earlier, use cipher suites which require ephemeral keying.

The following section(s) describe guidance for RADIUS server implementations and deployments. We first give an overview of current practices, and then extend and/or replace those practices for TLS-PSK.

Current Practices RADIUS identifies clients by source IP address ( and ) or by client certificate ( and ). Neither of these approaches work for TLS-PSK. This section describes current practices and mandates behavior for servers which use TLS-PSK. A RADIUS/UDP server is typically configured with a set of information per client, which includes at least the source IP address and shared secret. When the server receives a RADIUS/UDP packet, it looks up the source IP address, finds a client definition, and therefore the shared secret. The packet is then authenticated (or not) using that shared secret. That is, the IP address is treated as the clients identity, and the shared secret is used to prove the clients authenticity and shared trust. The set of clients forms a logical database "client table", with the IP address as the key. A server may be configured with additional site-local policies associated with that client. For example, a client may be marked up as being a WiFi Access Point, or a VPN concentrator, etc. Different clients may be permitted to send different kinds of requests, where some may send Accounting-Request packets, and other clients may not send accounting packets.

Practices for TLS-PSK We define practices for TLS-PSK by analogy with the RADIUS/UDP use-case, and by extending the additional policies associated with the client. The PSK identity replaces the source IP address as the client identifier. The PSK replaces the shared secret as proof of client authenticity and shared trust. However, systems implementing RADIUS/TLS and RADIUS/DTLS MUST still use the shared secret as discussed in those specifications. Any PSK is only used by the TLS layer, and has no effect on the RADIUS data which is being transported. That is, the RADIUS data transported in a TLS tunnel is the same no matter if client authentication is done via PSK or by client certificates. The encoding of the RADIUS data is entirely unaffected by the use (or not) of PSKs and client certificates. In order to securely support dynamic source IP addresses for clients, we also require that servers limit clients based on a network range. The alternative would be to suggest that RADIUS servers allow any source IP address to connect and try TLS-PSK, which could be a security risk. When RADIUS servers do no source IP address filtering, it is easier for attackers to send malicious traffic to the server. An issue with a TLS library or even a TCP/IP stack could permit the attacker to gain unwarranted access. In contrast, when IP address filtering is done, attackers generally must first gain access to a secure network before attacking the RADIUS server. Even where dynamic discovery is not used, servers SHOULD NOT permit TLS-PSK to be used across the wider Internet. The intent for TLS-PSK is for it to be used in internal / secured networks, where clients come from a small number of known locations. In contrast, certificates can be generated and assigned to clients without any interaction with the RADIUS server. Therefore if the RADIUS server needs to accept connections from clients at unknown locations, a more secure method is to use client certificates. If a client system is compromised, its complete configuration is exposed to the attacker. Exposing a client certificate means that the attacker can pretend to be the client. In contrast, exposing a PSK means that the attacker can not only pretend to be the client, but can also pretend to be the server. The benefits of TLS-PSK are in easing management and in administrative overhead, not in securing traffic from resourceful attackers. Where TLS-PSK is used across the Internet, PSKs MUST contain at least 256 bits of entropy. For example, a RADIUS server could be configured to be accept connections from a source network of 192.0.2.0/24. The server could therefore discard any TLS connection request which comes from a source IP address outside of that network. In that case, there is no need to examine the PSK identity or to find the client definition. Instead, the IP source filtering policy would deny the connection before any TLS communication had been performed. RADIUS servers need to be able to limit certain PSK identifiers to certain network ranges or IP addresses. That is, if a NAS is known to have a dynamic IP address within a particular subnet, the server should limit use of the NASes PSK to that subnet. This filtering can therefore help to catch configuration errors. As some clients may have dynamic IP addresses, it is possible for a one PSK identity to appear at different source IP addresses over time. In addition, as there may be many clients behind one NAT gateway, there may be multiple RADIUS clients using one public IP address. RADIUS servers need to support multiple PSK identifiers at one source IP address. That is, a server needs to support multiple different clients within one network range, multiple clients behind a NAT, and one client having different IP addresses over time. All of those use-cases are common and necessary. The following section describes these requirements in more detail.

IP Filtering A server supporting this specification MUST perform IP address filtering on incoming connections. There are few reasons for a server to have a default configuration which allows connections from any source IP address. A TLS-PSK server MUST be configurable with a set of "allowed" network ranges from which clients are permitted to connect. Any connection from outside of the allowed range(s) MUST be rejected before any PSK identity is checked. It is RECOMMENDED that servers support IP address filtering even when TLS-PSK is not used. The "allowed" network ranges could be implemented as a global list, or one or more network ranges could be tied to a client or clients. The intent here is to allow connections to be filtered by source IP, and to allow clients to be limited to a subset of network addresses. The exact method and representation of that filtering is up to an implementation. Conceptually, the set of IP addresses and ranges, along with permitted clients and their credentials forms a logical "client table" which the server uses to both filter and authenticate clients. The client table should contain information such as allowed network ranges, PSK identity and associated PSK, credentials for another TLS authentication method, or flags which indicate that the server should require a client certificate. Once a server receives a connection, it checks the source IP address against the list of all allowed IP addresses or ranges in the client table. If none match, the connection MUST be rejected. That is, the connection MUST be from an authorized source IP address. Once a connection has been established, the server MUST NOT process any application data inside of the TLS tunnel until the client has been authenticated. Instead, the server normally receives a TLS-PSK identity from the client. The server then uses this identity to look up the client in the client table. If there is no matching client, the server MUST close the connection. The server then also checks if this client definition allows this particular source IP address. If the source IP address is not allowed, the server MUST close the connection. Where the server does not receive TLS-PSK from the client, it proceeds with another authentication method such as client certificates. Such requirements are discussed elsewhere, most notably in and . Returning to the subject of IP address lookups, an implementation may perform two independent IP address lookups. First, to check if the connection allowed at all, and second to check if the connection is authorized for this particular client. One or both checks may be used by a particular implementation. The two sets of IP addresses can overlap, and implementations SHOULD support that capability. Depending on the implementation, one or more clients may share a list of allowed network ranges. Alternately, the allowed network ranges for two clients can overlap only partially, or not at all. All of these possibilities MUST be supported by the server implementation.

PSK Authentication Once the source IP has been verified to be allowed for this particular client, the server authenticates the TLS connection via the PSK taken from the client definition. If the PSK is verified, the server then accepts the connection, and proceeds with RADIUS/TLS as per . If the PSK is not verified, then the server MUST close the connection. While TLS provides for fallback to other authentication methods such as client certificates, there is no reason for a client to be configured simultaneously with multiple authentication methods. A client MUST use only one authentication method for TLS. An authentication method is either TLS-PSK, client certificates, or some other method supported by TLS. That is, client configuration is relatively simple: use a particular set of credentials to authenticate to a particular server. While clients may support multiple servers and fail-over or load-balancing, that configuration is generally orthogonal to the choice of which credentials to use.

Resumption Implementations SHOULD support resumption. In many cases session tickets can be authenticated solely by the server, and do not require querying a database. The use of resumption can allow the system to better scale to higher loads. However, the above discussion of PSK identities is complicated by the use of PSKs for resumption in TLS 1.3. A server which receives a PSK identity via TLS typically cannot query the TLS layer to see if this identity is for a resumed session, or is instead a static pre-provisioned identity. This confusion complicates server implementations. One way for a server to tell the difference between the two kinds of identities is via construction. Identities used for resumption can be constructed via a fixed format, such as recommended by . A static pre-provisioned identity could be in format of an NAI, as given in . An implementation could therefore examine the incoming identity, and determine from the identity alone what kind of authentication was being performed. An alternative way for a server to distinguish the two kinds of identities is to maintain two tables. One table would contain static identities, as the logical client table described above. Another table could be the table of identities handed out for resumption. The server would then look up any PSK identity in one table, and if not found, query the other one. An identity would be found in a table, in which case it can be authenticated. Or, the identity would not be found in either table, in which case it is unknown, and the server MUST close the connection. As suggested in , TLS-PSK peers MUST NOT store resumption PSKs or tickets (and associated cached data) for longer than 604800 seconds (7 days) regardless of the PSK or ticket lifetime. Systems supporting TLS-PSK and resumption MUST cache data during the initial full handshake sufficient to allow authorization decisions to be made during resumption. If cached data cannot be retrieved securely, resumption MUST NOT be done. The cached data is typically information such as the original PSK identity, along with any policies associated with that identity. Information from the original TLS exchange (e.g., the original PSK identity) as well as related information (e.g., source IP addresses) may change between the initial full handshake and resumption. This change creates a "time-of-check time-of-use" (TOCTOU) security vulnerability. A malicious or compromised client could supply one set of data during the initial authentication, and a different set of data during resumption, potentially allowing them to obtain access that they should not have. If any authorization or policy decisions were made with information that has changed between the initial full handshake and resumption, and if change may lead to a different decision, such decisions MUST be reevaluated. It is RECOMMENDED that authorization and policy decisions are reevaluated based on the information given in the resumption. TLS-PSK servers MAY reject resumption where the information supplied during resumption does not match the information supplied during the original authentication. If a safe decision is not possible, TLS-PSK servers SHOULD reject the resumption and continue with a full handshake.

Interaction with other TLS authentication methods When a server supports both TLS-PSK and client certificates, it MUST be able to accept authenticated connections from clients which may use either type of credentials, perhaps even from the same source IP address and at the same time. That is, servers are required to both authenticate the client, and also to filter clients by source IP address. These checks both have to match in order for a client to be accepted.

Privacy Considerations We make no changes over and .

Security Considerations The primary focus of this document is addressing security considerations for RADIUS.

IANA Considerations There are no IANA considerations in this document. RFC Editor: This section may be removed before final publication.

Acknowledgments Thanks to the many reviewers in the RADEXT working group for positive feedback.

Changelog

00 - initial version
01 - update examples