Agent Authorization Profile (AAP) for OAuth 2.0

Introduction The Agent Authorization Profile (AAP) is an authorization profile built on OAuth 2.0 and JWT , designed to support secure, auditable, and context-aware authorization for autonomous AI agents. AAP extends existing standards with structured claims and validation rules that allow systems to reason about agent identity, task context, operational constraints, delegation chains, and human oversight requirements.

Problem Statement Traditional OAuth-based systems were designed primarily for user-to-application-to-API interactions. Autonomous AI agents introduce different characteristics: actions can be autonomous and high frequency; human approval may not be present at execution time; risk depends on task context; and execution may be delegated across multiple tools or agents. Existing scope-based models are not expressive enough to represent these requirements safely and transparently.

Goals

Provide explicit and verifiable identity for AI agents.
Support capability-based authorization with enforceable constraints.
Bind access tokens to specific tasks and declared purposes.
Enable auditable delegation across agents and tools.
Support the expression of human oversight requirements.
Remain compatible with OAuth 2.0, JWT, Token Exchange, and proof-of-possession mechanisms.

Non-Goals

Defining internal AI model behavior.
Judging the correctness or ethics of agent decisions.
Replacing organizational security or compliance frameworks.
Standardizing log storage formats or SIEM integrations.

Terminology

Term	Definition

Agent	Autonomous software entity (e.g. LLM, bot) that acts as an OAuth client and performs actions on behalf of an operator.
Authorization Server (AS)	Server that issues access tokens in accordance with OAuth 2.0 and applies AAP policies.
Capability	Permitted action (e.g. `action`) together with its restrictions (`constraints`).
Delegation	Transfer of a subset of privileges to another entity (tool or sub-agent) via token exchange or other mechanism.
Operator	Organization or human role that registers and authorizes the agent.
Proof-of-possession (PoP)	Mechanism by which the client demonstrates possession of a key (e.g. DPoP, mTLS) when using the token.
Resource Server (RS)	Server that protects resources and validates AAP tokens before allowing access.
Task	Unit of work to which the token is bound (identifier, purpose, sensitivity, etc.).

AAP token: An access token issued by an Authorization Server that conforms to this profile and contains AAP claims (e.g. aap_agent, aap_task, aap_capabilities). Claim: A name/value pair in a JWT payload. AAP defines additional claim names and structures for agent identity, task binding, capabilities, oversight, delegation, context, and audit.

Conventions Used in This Document 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. For conformance: "MUST" and "SHALL" indicate mandatory requirements; "SHOULD" and "RECOMMENDED" indicate recommended but not mandatory behavior; "MAY" and "OPTIONAL" indicate optional behavior.

Overview and Relationship to Existing Standards AAP operates within a standard OAuth architecture consisting of an Authorization Server (AS), Resource Servers (RS), and clients. In AAP, the client is an autonomous AI agent. Token issuance follows OAuth 2.0; typically the Client Credentials Grant Section 4.4 is used for agent-to-API (M2M) flows. Client (agent) authentication MAY use standard client authentication (client secret, mTLS, etc.) or assertions (e.g. JWT-based client authentication) as per the deployment profile. When the agent is a workload identified by SPIFFE , the AS MAY accept SVIDs or derived tokens as part of client authentication; AAP does not define a new flow but MAY integrate with SPIFFE. Tokens issued by the AS include additional structured claims that Resource Servers MUST evaluate before allowing operations. AAP does not introduce a new identity or protocol scheme; it reuses existing standards and adds a layer of claims and validation rules.

OAuth 2.0 — AAP uses the standard OAuth flow (AS, RS, client); the client is the agent; tokens are JWTs with additional AAP claims.
OpenID Connect — Agent identity MAY be based on OIDC sub and iss; AAP adds agent-, task-, and capability-specific claims.
mTLS — RECOMMENDED for proof-of-possession and for agent authentication toward the AS and RS.
DPoP — Alternative to mTLS for proof-of-possession; RECOMMENDED when mTLS is not feasible.
SPIFFE — OPTIONAL; the agent identifier (e.g. in aap_agent) MAY be a SPIFFE ID (spiffe://trust-domain/...) when the deployment uses SPIFFE/SPIRE for workload identity.
Token Exchange — Used for delegation and for privilege reduction on token re-issuance; the act (actor) claim MAY be used in the delegation chain.

JWT Claim Schema (AAP Profile) AAP tokens extend standard JWT claims with the following structured sections. To avoid collisions, AAP uses a registered claim namespace; the normative claim names are: aap_agent, aap_task, aap_capabilities, aap_oversight, aap_delegation, aap_context, aap_audit. Formal Schema: Complete JSON Schema definitions for all AAP claims are provided in Appendix A and in the /schemas directory of the reference implementation. Implementations SHOULD validate tokens against these schemas to ensure conformance.

Claim Semantics

Agent identity — MAY be expressed via OIDC sub and iss, or via the aap_agent claim. The aap_agent claim MAY contain, among other fields, a SPIFFE ID (spiffe://trust-domain/...) when the deployment uses SPIFFE/SPIRE for workload identity.
Delegation — The delegation chain MAY use the standard act (actor) claim from . Optionally, aap_delegation MAY carry additional metadata (e.g. depth, origin) when more than act is required.
Oversight — Human oversight requirements are expressed as policy metadata in aap_oversight (e.g. requires_approval_for for certain capability types, or max_autonomous_scope). AAP only carries the intent; enforcement is at the Resource Server or orchestrator (e.g. OIDC step-up with acr_values or an external approval API), and is out of scope for this profile.
Audit — Trace identifiers in aap_audit SHOULD be compatible with existing trace context propagation (e.g. W3C Trace Context, OpenTelemetry ) so that logs can be correlated with distributed traces without defining a new audit schema.

Structured Sections (Claim Names)

aap_agent
aap_task
aap_capabilities
aap_oversight
aap_delegation (and/or act per )
aap_context
aap_audit

Example Claim Structures The following examples illustrate concrete claim shapes. In the profile, these may appear under the normative names above (e.g. aap_agent / agent). Standard JWT claims (iss, sub, aud, exp, iat, jti) are assumed. Agent identity — Identifies the autonomous agent and its execution context. json { "agent": { "id": "agent-researcher-01", "type": "llm-autonomous", "operator": "org:blogcorp", "model": { "provider": "provider-name", "id": "model-id", "version": "model-version" }, "runtime": { "environment": "kubernetes", "attested": true } } } Task binding — Binds the token to a specific task and purpose. json { "task": { "id": "task-id", "purpose": "research_and_draft_article", "topic": "example-topic", "data_sensitivity": "public", "created_by": "user-id" } } Capabilities — Authorized actions and constraints. json { "capabilities": [ { "action": "search.web", "constraints": { "domains_allowed": ["example.org"], "max_requests_per_hour": 50 } }, { "action": "cms.create_draft", "constraints": { "status": "draft_only" } } ] } Oversight — Human approval requirements for certain actions. json { "oversight": { "requires_human_approval_for": ["cms.publish", "execute.payment"], "approval_reference": "policy-id" } } Delegation — Depth and chain of delegation. json { "delegation": { "depth": 0, "max_depth": 2, "chain": ["agent:agent-researcher-01"] } } Context — Network, time, and geo restrictions. json { "context": { "network_zone": "public-internet-only", "time_window": { "start": "ISO-8601", "end": "ISO-8601" }, "geo_restriction": "none" } } Audit — Trace and session identifiers for logging. json { "audit": { "log_level": "full", "trace_id": "trace-id", "session_id": "session-id" } }

Complete Example Payload The following is a single JSON object representing the decoded payload of an AAP access token (claims only; signature and encoding are per RFC 7519). Standard JWT claims and all AAP claims are combined with consistent values (e.g. same agent.id and task.id referenced in audit). json { "iss": "https://as.example.com", "sub": "agent-researcher-01", "aud": "https://api.example.com", "exp": 1704067200, "iat": 1704063600, "jti": "token-unique-id-123", "agent": { "id": "agent-researcher-01", "type": "llm-autonomous", "operator": "org:blogcorp", "model": { "provider": "provider-name", "id": "model-id", "version": "model-version" }, "runtime": { "environment": "kubernetes", "attested": true } }, "task": { "id": "task-id", "purpose": "research_and_draft_article", "topic": "example-topic", "data_sensitivity": "public", "created_by": "user-id" }, "capabilities": [ { "action": "search.web", "constraints": { "domains_allowed": ["example.org"], "max_requests_per_hour": 50 } }, { "action": "cms.create_draft", "constraints": { "status": "draft_only" } } ], "oversight": { "requires_human_approval_for": ["cms.publish", "execute.payment"], "approval_reference": "policy-id" }, "delegation": { "depth": 0, "max_depth": 2, "chain": ["agent:agent-researcher-01"] }, "context": { "network_zone": "public-internet-only", "time_window": { "start": "2024-01-01T00:00:00Z", "end": "2024-01-01T23:59:59Z" }, "geo_restriction": "none" }, "audit": { "log_level": "full", "trace_id": "trace-id", "session_id": "session-id" } }

Action Name Grammar (ABNF) Action names in the action field of capabilities MUST conform to the following ABNF grammar : abnf action-name = component *( "." component ) component = ALPHA *( ALPHA / DIGIT / "-" / "_" ) Where: - ALPHA is any ASCII alphabetic character (a-z, A-Z) - DIGIT is any ASCII digit (0-9) - Component names MUST start with an alphabetic character - Component names MAY contain hyphens and underscores after the first character - Action names are formed by concatenating components with dots (.) Examples of valid action names: - search.web - cms.create_draft - cms.publish - execute.payment - api.v2.users.read - data-pipeline.transform_records Examples of invalid action names: - search..web (double dot) - .search.web (starts with dot) - search.web. (ends with dot) - 9api.read (component starts with digit) - search.web* (wildcard not allowed) Matching Semantics: Resource Servers MUST perform exact string matching on action names. Wildcard matching (e.g., cms.* matching cms.publish) is NOT part of this specification but MAY be defined in future extensions. Action names are case-sensitive. search.Web and search.web are different actions. Versioning MAY be expressed through namespace components: - api.v1.search.web - api.v2.search.web

Standard Constraint Types and Semantics This section defines the semantics of standard constraint types used within capability constraints. Resource Servers MUST enforce these constraints according to the semantics defined here. Constraint Enforcement Semantics: When multiple capabilities match the same action: - OR semantics: If ANY capability grants the action, the request is authorized (subject to that capability's constraints) - Resource Server evaluates capabilities in order and uses the first match When multiple constraints exist within a single capability: - AND semantics: ALL constraints MUST be satisfied for the action to be authorized - If any constraint fails, the entire request MUST be denied

Rate Limiting Constraints

Constraint Name	Type	Semantics	Example
`max_requests_per_hour`	integer	Fixed hourly quota. Window resets at minute 0 of each hour (clock hour). Failed requests count toward quota. Retries count as new requests.	`50`
`max_requests_per_minute`	integer	Sliding 60-second window from current request time backwards. Resource Server MUST track request timestamps.	`10`
`max_requests_per_day`	integer	Fixed daily quota. Window resets at 00:00:00 UTC. Failed requests count toward quota.	`1000`

Implementation Notes: - Rate limits are per token (identified by jti claim) - Resource Servers SHOULD use distributed rate limiting for multi-instance deployments - On quota exceeded: Resource Server MUST return HTTP 429 with aap_constraint_violation error - Rate limit state SHOULD be cleared when token expires

Domain and Network Constraints

Constraint Name	Type	Semantics	Example
`domains_allowed`	array of strings	DNS suffix matching (rightmost matching). `subdomain.example.org` matches `example.org` in allowlist. Resource Server MUST extract domain from request target URL and validate.	`["example.org", "trusted.com"]`
`domains_blocked`	array of strings	Blocklist takes precedence over allowlist. If both are present, blocked domains MUST be checked first.	`["malicious.com"]`
`ip_ranges_allowed`	array of CIDR strings	IP ranges in CIDR notation. Resource Server validates destination IP of request.	`["192.168.1.0/24"]`

Domain Matching Algorithm: 1. Extract domain from request target URL 2. If domains_blocked is present and domain matches any blocked entry: DENY 3. If domains_allowed is present: a. Check if domain is exact match or has allowed domain as suffix b. If match found: ALLOW (proceed to other constraints) c. If no match: DENY 4. If neither constraint present: proceed to other constraints

Time-Based Constraints

Constraint Name	Type	Semantics	Example
`time_window.start`	ISO 8601 string	Request timestamp MUST be after or equal to this time (inclusive). Resource Server uses its own clock with max 5-minute skew tolerance.	`"2024-01-01T00:00:00Z"`
`time_window.end`	ISO 8601 string	Request timestamp MUST be before this time (exclusive).	`"2024-12-31T23:59:59Z"`

Clock Skew Handling: - Resource Server clock is authoritative - Resource Server MAY tolerate up to 5 minutes of clock skew - If request timestamp is outside window beyond skew tolerance: DENY with aap_constraint_violation

Delegation Constraints

Constraint Name	Type	Semantics	Example
`max_depth`	integer (0-10)	Maximum delegation depth for this capability. 0 means no delegation allowed. Resource Server MUST validate `delegation.depth <= max_depth`.	`2`

Data and Security Constraints

Constraint Name	Type	Semantics	Example
`max_response_size`	integer	Maximum response size in bytes. Resource Server SHOULD enforce during response streaming.	`10485760` (10MB)
`max_request_size`	integer	Maximum request payload size in bytes. Resource Server MUST validate before processing.	`1048576` (1MB)
`data_classification_max`	enum string	Maximum data classification level accessible. Values: `public`, `internal`, `confidential`, `restricted`. Resource Server enforces based on resource classification.	`"internal"`
`allowed_methods`	array of strings	HTTP methods allowed. Resource Server MUST validate request method against this list.	`["GET", "POST"]`
`allowed_regions`	array of ISO 3166-1 alpha-2 codes	Geographic regions where requests are allowed. Resource Server validates based on request origin or target resource location.	`["US", "CA", "GB"]`

Delegation Chain Semantics The delegation claim tracks authorization delegation across agents and tools using OAuth Token Exchange . Delegation Depth Calculation: depth = 0: Original agent token (no delegation) depth = 1: Token obtained via Token Exchange from depth=0 token depth = n: Token obtained via Token Exchange from depth=n-1 token Authorization Server Requirements: - AS MUST increment delegation.depth by 1 on each Token Exchange - AS MUST append the current agent/tool identifier to delegation.chain array - AS MUST copy and preserve delegation.chain from parent token - AS MUST NOT issue token if resulting depth would exceed capability's max_depth constraint - AS MUST NOT issue token if resulting depth exceeds delegation.max_depth claim Resource Server Requirements: - RS MUST reject requests if delegation.depth > delegation.max_depth - RS MUST validate delegation depth against capability-specific max_depth constraints - RS MUST validate that delegation.chain length equals delegation.depth + 1 Delegation Chain Format: json "delegation": { "depth": 2, "max_depth": 3, "chain": [ "spiffe://trust.example.com/agent/researcher-01", // depth=0 (origin) "spiffe://trust.example.com/tool/web-scraper", // depth=1 "https://as.example.com/agents/translator" // depth=2 (current) ], "parent_jti": "parent-token-jti-value" } Privilege Reduction Requirements: When issuing a derived token via Token Exchange, the Authorization Server MUST reduce privileges by one or more of: - Removing capabilities (subset of parent capabilities) - Adding stricter constraints (lower rate limits, narrower domain lists) - Reducing token lifetime (shorter exp time) - Reducing max_depth (limit further delegation) The Authorization Server MUST NOT grant capabilities not present in the parent token. Preventing Confused Deputy Attacks: To prevent confused deputy attacks where a delegated token is replayed: - Each token MUST have a unique jti (JWT ID) - Delegation chain MUST be immutable (copied, never modified) - Token Exchange MUST record parent_jti linking to parent token - Authorization Server MAY implement token family revocation (revoking parent revokes all descendants)

Threat Model Summary AAP assumes environments where autonomous AI agents can access APIs, perform chained actions, and operate for extended periods without direct human intervention. The following threats are in scope; for each, agent-specific risks and AAP mitigations are noted.

Agent Impersonation Threat: An attacker obtains agent credentials or steals a token and acts as an authorized agent. Agent-specific risk: An agent may have broad permissions and act many times per minute, amplifying impact. Mitigations: Short-lived tokens; Proof-of-Possession (mTLS or DPoP); attested workload identity when possible; strong agent identity claims (aap_agent / agent.id, agent.model, runtime.attested).

Capability Escalation Threat: The agent attempts actions beyond what is authorized (e.g. publish instead of create draft). Agent-specific risk: Agents may generate new strategies or calls not anticipated by the developer. Mitigations: Structured capabilities with constraints (not broad scopes); mandatory validation of action + constraints by the Resource Server; task-bound tokens (task.purpose); explicit separation between automatic and human-supervised actions.

Purpose Drift Threat: A token issued for one task is reused for another (e.g. token for “public health research” used for “extracting sensitive data”). Mitigations: Mandatory task claim with purpose; Resource Servers verify consistency between declared purpose and requested operation; short time windows; reject requests that do not match the declared context.

Malicious or Excessive Delegation Threat: An agent delegates to tools or sub-agents with more privileges than intended. Agent-specific risk: Agent ecosystems are often modular and chained. Mitigations: OAuth Token Exchange with privilege reduction; delegation.depth and delegation.chain claims; maximum depth limit (max_depth); prohibition of delegation for certain critical capabilities.

Large-Scale Automated Misuse Threat: An authorized agent performs valid actions at harmful volume (spam, abusive scraping, mass resource creation). Mitigations: Quantitative constraints in capabilities (max_requests_per_hour, etc.); enforced by the Resource Server; monitoring and rapid token revocation; mandatory audit with traceability by task and agent.

Prompt / Data Injection Threat: A third party manipulates external data to induce the agent to use its permissions in unwanted ways. Note: AAP does not control the AI model but can limit impact. Mitigations: Tokens bound to specific purpose; restriction of domains, action types, and volumes; separation of read vs. write vs. execute capabilities; human oversight required for high-impact actions.

Lack of Traceability Threat: Inability to reconstruct which agent did which action under which authorization. Agent-specific risk: Decisions can be complex and chained. Mitigations: Audit claims (audit.trace_id, task.id); mandatory propagation of trace identifiers; inclusion of delegation chain in derived tokens.

Use Outside Intended Environment Threat: A valid token is used from a network, region, or environment other than the one authorized. Mitigations: Context claims (context.network_zone, time windows); additional validation by the Resource Server; combination with traditional network controls.

Summary AAP assumes that agents are potentially powerful and highly automated; risk depends not only on who accesses but on purpose, limits, and delegation chain; authorization MUST be contextual, restricted, and auditable. AAP extends OAuth from a broad-permission model toward verifiable operational contracts between organizations, agents, and services.

Resource Server Validation Rules This section defines the validation rules that a Resource Server (RS) MUST apply before accepting a request authenticated with an AAP access token. These rules extend standard OAuth 2.x token validation with agent-specific, task-bound, and capability-aware checks.

Standard Token Validation

The RS MUST verify the token signature using trusted Authorization Server keys.
The RS MUST verify the token has not expired (exp claim) and is within acceptable clock skew.
The RS MUST verify the audience (aud) matches the Resource Server.
The RS MUST verify the issuer (iss) is trusted.
The RS MUST verify the token has not been revoked if a revocation or introspection mechanism is in place (, when introspection is used).

Proof of Possession Validation For AAP tokens, proof-of-possession is RECOMMENDED; for high-risk profiles it SHOULD be REQUIRED. Implementations MUST support at least one of: DPoP or mTLS client authentication . If mTLS or DPoP is used, the RS MUST validate that the requester demonstrates possession of the key bound to the token. Bearer-only usage is not sufficient for high-risk agent capabilities.

Agent Identity Validation

The RS MUST ensure the agent (or aap_agent) claim is present and well-formed.
The RS MUST verify agent.id is recognized or allowed by local policy.
If present, the RS MUST evaluate agent.runtime.attested according to local trust requirements.
If model information is included, the RS MUST ensure the model identifier is not on a deny list.

Task Binding Validation

The RS MUST ensure the task (or aap_task) claim is present for agent-issued tokens.
The RS MUST verify the current request is consistent with task.purpose.
The RS MUST reject requests that clearly fall outside the declared purpose or data sensitivity.
The RS MAY enforce that the token is used only within the declared time window.

Capability Enforcement The Resource Server MUST treat the capabilities (or aap_capabilities) claim as the authoritative source of permitted actions.

The RS MUST match the requested operation to a capability.action entry.
The RS MUST enforce all constraints associated with the matching capability.
The RS MUST deny the request if no matching capability is found.
The RS MUST apply quantitative limits such as rate limits or volume caps defined in constraints.

Oversight Requirement Enforcement

If the requested action appears in oversight.requires_human_approval_for (or equivalent in aap_oversight), the RS MUST NOT complete the action automatically.
The RS SHOULD return a response indicating that human approval is required.
The RS MAY provide a reference to the approval workflow indicated by approval_reference.

Delegation Chain Validation

If a delegation (or aap_delegation) claim is present, the RS MUST verify that delegation.depth does not exceed local policy limits.
The RS MUST inspect delegation.chain to understand upstream actors.
The RS SHOULD apply stricter policy if the chain includes untrusted or unknown actors.
The RS MUST ensure delegated tokens do not contain broader capabilities than the original agent token.

Contextual Restrictions Enforcement

If context.network_zone is present, the RS MUST verify the request originates from an allowed environment when technically feasible.
The RS MUST enforce time window constraints if context.time_window is present.
The RS MUST apply additional checks for geo or network restrictions when provided.

Audit and Trace Propagation

The RS MUST extract audit.trace_id and propagate it to internal logs.
The RS MUST log agent.id, task.id, action performed, and authorization decision outcome.
The RS MUST ensure logs are protected against tampering according to organizational policy.

Failure Handling

If any mandatory validation step fails, the RS MUST deny the request.
Error responses SHOULD avoid leaking sensitive authorization details.
Repeated violations MAY trigger rate limiting or temporary blocking of the agent identity.

Error responses: On validation failure, the RS SHOULD respond with HTTP 403 Forbidden when the token is valid but the request is not authorized (e.g. capability mismatch, task inconsistency, oversight required). Use HTTP 401 Unauthorized when the token is invalid, expired, or missing (per OAuth 2.0 practice, e.g. ). The response body SHOULD follow a structure such as error and optional error_description (e.g. RFC 6749 / RFC 6750 style) without revealing internal authorization details—e.g. use generic error codes such as insufficient_scope or invalid_request rather than describing the specific rule that failed (e.g. avoid “capability X not in token”). Avoid including in error_description the exact authorization rule that failed, so as not to leak information to an attacker.

Authorization Server Requirements This section defines the requirements that an Authorization Server (AS) MUST satisfy to issue AAP access tokens. These extend standard OAuth 2.0 token issuance with agent-specific binding, capabilities, and audit.

Authentication

Strongly authenticate agents before issuing tokens (e.g. Client Credentials Grant per RFC 6749 Section 4.4).
Support client authentication via client secret, mTLS, or assertions (e.g. JWT-based client authentication) as per the deployment profile.
Optionally integrate with SPIFFE SVIDs when the agent is a workload identified by SPIFFE .

Token Binding

Bind issued tokens to a specific task and declared purpose.
Include task identifier and purpose in the token so that Resource Servers can verify request consistency.

Capabilities and Constraints

Embed in the token only capabilities and constraints authorized by AS policy for the agent and task.
Do not issue tokens with broader capabilities than the operator has authorized for the given task.

Token Lifetime

Limit token lifetimes according to risk level (e.g. shorter for high-impact or high-frequency use).
Apply policy-based rules for expiration and acceptable clock skew.

Token Exchange

Support token exchange per when delegation or privilege reduction is required.
Enforce reduction of privileges when mapping subject_token to issued_token (e.g. subset of aap_capabilities); mapping rules are AS policy.
Use the act (actor) claim when building delegation chains as per RFC 8693.

Revocation

Support rapid revocation per (Token Revocation) and/or RFC 7662 (Introspection) as appropriate to the deployment.
Ensure Resource Servers can check revocation status when policy requires it (e.g. via introspection endpoint or revocation list).

Proof-of-Possession

For AAP tokens, support at least one of DPoP or mTLS when the profile requires proof-of-possession.
Issue tokens that indicate PoP binding when the client has demonstrated key possession during the token request.

Audit

Record issuance events (agent identity, task, capabilities granted, timestamp) for audit and traceability.
Support correlation with trace identifiers included in the token (aap_audit / audit) where applicable.

Example High-Level Flow

An operator defines allowed capabilities and policies for an agent.
The agent authenticates with the Authorization Server (e.g. Client Credentials Grant).
The agent requests a task-bound access token; optional token exchange (RFC 8693) for delegation or privilege reduction.
The Authorization Server issues a JWT containing AAP claims.
The agent calls Resource Servers using the token (with DPoP or mTLS when proof-of-possession is required).
Resource Servers validate the token and enforce capabilities and constraints.
All actions are logged with trace identifiers for audit.

Extensibility AAP is designed as a profile and allows additional claims or constraints to be defined by industry groups or organizations, provided they do not weaken core validation requirements.

Conformance A conforming implementation satisfies the requirements of this profile for its role (Authorization Server or Resource Server).

Authorization Server: A conforming AS MUST satisfy all requirements in Section 8 (Authorization Server Requirements). It MUST issue tokens that include the AAP claims required by the deployment profile and MUST support proof-of-possession (DPoP and/or mTLS) when the profile requires it. It SHOULD support token exchange and revocation RFC7009 as appropriate.
Resource Server: A conforming RS MUST apply all rules in Section 7 (Resource Server Validation Rules) (standard token validation, proof-of-possession when required, agent identity, task binding, capability enforcement, oversight, delegation chain, contextual restrictions, audit and trace propagation, failure handling). It MUST deny requests when any mandatory validation step fails and MUST NOT leak sensitive authorization details in error responses.
A conforming implementation MAY support additional claims or options provided they do not weaken the requirements above.

Security Considerations This section addresses security considerations specific to AAP beyond those covered in OAuth 2.0 Security Best Current Practice and related standards.

Cryptographic Algorithms and Key Management Token Signing (REQUIRED): Authorization Servers MUST sign AAP tokens using asymmetric cryptography. The following algorithms are REQUIRED:

ES256 (ECDSA with P-256 curve and SHA-256): RECOMMENDED for new deployments
- Provides 128-bit security level
- Smaller signatures than RSA
- Fast signature verification
RS256 (RSA with SHA-256): REQUIRED for backward compatibility
- Minimum 2048-bit RSA keys
- 3072-bit or 4096-bit RSA keys RECOMMENDED for long-lived keys or high-security environments

Authorization Servers MUST NOT use symmetric algorithms (HS256) for production AAP tokens, as symmetric keys require sharing between AS and RS, which increases key exposure risk. Proof-of-Possession Algorithms: When using DPoP : - Agents MUST use ES256 for DPoP proof generation - RS256 MAY be supported for legacy clients - DPoP proof lifetime SHOULD be short (maximum 60 seconds) When using mTLS : - TLS 1.3 REQUIRED; TLS 1.2 MAY be supported with restricted cipher suites - Cipher suites: ECDHE_ECDSA or ECDHE_RSA with AES_GCM - Client certificates MUST be validated against trusted CA or certificate pinning Key Sizes:

Key Type	Minimum	Recommended	Security Level
RSA	2048-bit	3072-bit	~112-bit / ~128-bit
ECDSA	P-256	P-384	128-bit / 192-bit
Symmetric (client secrets)	128-bit entropy	256-bit entropy	128-bit / 256-bit

Key Rotation:

AS signing keys SHOULD be rotated every 90 days
Previous keys MUST be retained for token validation during overlap period (RECOMMENDED: 24 hours after rotation)
Resource Servers MUST support validation with multiple concurrent AS public keys
Key rotation MUST be coordinated via JWKS (JSON Web Key Set) endpoint

Key Storage:

Authorization Server private signing keys MUST be stored in Hardware Security Modules (HSM) or equivalent secure key management services for production deployments
Agent client credentials (secrets, private keys) SHOULD be stored in secure vaults (e.g., HashiCorp Vault, cloud KMS)
Keys MUST NOT be logged, included in error messages, or transmitted over insecure channels

Proof-of-Possession Requirements Risk Assessment: Bearer tokens (tokens without proof-of-possession) present elevated risk for autonomous agents due to: - High request rates amplify impact of stolen tokens - Agents may operate unattended for extended periods - Token theft from agent memory or logs enables replay attacks Deployment Profiles: AAP defines three security profiles for proof-of-possession:

Profile	PoP Requirement	Use Case
Strict	REQUIRED (mTLS or DPoP)	Production systems, confidential data, high-risk capabilities
Standard	RECOMMENDED	Development, internal tools, low-risk capabilities
Legacy	OPTIONAL	Migration scenarios, backward compatibility

Profile Selection Criteria: Proof-of-Possession SHOULD be REQUIRED when any of the following apply: - Capability includes data_classification_max >= "confidential" - Capability includes write, delete, or execute actions - oversight.level >= "approval" - Token lifetime > 1 hour - Agent accesses resources in different trust domains Token Binding (cnf claim): When PoP is used, tokens MUST include the cnf (confirmation) claim : For DPoP: json "cnf": { "jkt": "0ZcOCORZNYy-DWpqq30jZyJGHTN0d2HglBV3uiguA4I" } For mTLS: json "cnf": { "x5t#S256": "bwcK0esc3ACC3DB2Y5_lESsXE8o9ltc05O89jdN-dg" } Resource Servers MUST validate the proof matches the cnf claim before accepting the token.

Token Lifetime and Revocation Token Lifetime Guidelines:

Capability Risk Level	Recommended Lifetime	Maximum Lifetime
Read-only, public data	60 minutes	120 minutes
Read-only, internal data	30 minutes	60 minutes
Write, low-risk	15 minutes	30 minutes
Write, high-risk	5 minutes	15 minutes
Execute, delete	5 minutes	10 minutes

Token lifetime SHOULD be reduced for: - Derived tokens (via Token Exchange): 50% of parent lifetime - Tokens with high delegation depth: Reduce by 25% per depth level - Tokens without proof-of-possession: 50% of normal lifetime Revocation Requirements: Authorization Servers MUST support token revocation via one or both of: - Revocation endpoint : Agent or operator can revoke token - Token introspection : RS queries AS for token validity Rapid Revocation: For the purposes of this specification, "rapid revocation" is defined as: - Maximum 60 seconds from revocation request to enforcement by all Resource Servers - Authorization Server MUST distribute revocation events to Resource Servers within 30 seconds - Resource Server MUST apply revocation updates within 30 seconds of receipt Revocation Distribution: For multi-RS deployments, AS SHOULD use: - Push-based revocation (AS pushes to RS) rather than pull-based (RS polls AS) - Revocation event stream (e.g., Server-Sent Events, WebSockets, message queue) - Fallback to token introspection if revocation list fails to propagate Token Family Revocation: When a token is revoked: - Authorization Server MAY revoke all descendant tokens (derived via Token Exchange) - This is achieved by tracking parent_jti linkage - Token family revocation enhances security but requires AS to maintain token graph

Constraint Enforcement Server-Side Enforcement (REQUIRED): All capability constraints MUST be enforced by the Resource Server, NOT trusted to agent logic. Agents are potentially adversarial; they may attempt to bypass constraints. Rate Limiting in Distributed Systems: For Resource Servers deployed across multiple instances: - Rate limits MUST be enforced consistently across all instances - Use distributed rate limiting (e.g., Redis, shared counter service) - Accept eventual consistency with conservative limits (deny when uncertain) Constraint Validation Failures: When a constraint is violated, Resource Server MUST: - Return HTTP 403 (Forbidden) or 429 (Too Many Requests) as appropriate - Include error code aap_constraint_violation in response - Log violation event with agent.id, task.id, violated constraint - NOT return details of constraint values in error response (privacy)

Delegation Security Privilege Reduction (REQUIRED): When issuing derived tokens via Token Exchange , the Authorization Server MUST reduce privileges by one or more of: - Capability removal: Subset of parent capabilities only - Constraint tightening: Lower rate limits, narrower domain lists, shorter time windows - Lifetime reduction: Shorter exp time than parent token - Depth limit reduction: Lower max_depth to limit further delegation The Authorization Server MUST NOT grant capabilities not present in the parent token. "Privilege escalation via delegation" MUST be prevented. Delegation Depth Enforcement: Both Authorization Server and Resource Server MUST enforce delegation depth limits: - AS MUST NOT issue token if resulting delegation.depth would exceed delegation.max_depth - RS MUST reject requests if delegation.depth > delegation.max_depth - Defense-in-depth: Both layers validate to prevent bypass Confused Deputy Prevention: To prevent confused deputy attacks where delegation chains are replayed: - Each token MUST have unique jti (JWT ID) - Derived tokens MUST include delegation.parent_jti linking to parent token - AS MUST validate parent token exists and is not expired/revoked before issuing derived token - Delegation chain MUST be immutable (copied and appended, never modified by client)

Human Oversight Enforcement Approval Workflow Security: When oversight.requires_human_approval_for includes an action: - Resource Server MUST NOT execute action automatically - RS MUST return HTTP 403 with error code aap_approval_required - Response SHOULD include approval_reference URL for requesting approval - Approval mechanism is out of scope for AAP but may use OIDC step-up authentication, external workflow systems, etc. Approval Bypass Prevention:

Approval requirement is in token (signed by AS), not in request (untrusted)
Agent cannot bypass approval by modifying request headers
Resource Server MUST validate oversight claim before execution

Authorization Server Security AS as Critical Component: The Authorization Server is a critical security component. Compromise of the AS enables arbitrary token minting. AS Hardening Requirements:

Minimal software installation (reduce attack surface)
Regular security patching (OS, libraries, dependencies)
Network isolation (AS should not be internet-accessible if possible; use API gateway)
Access control (admin access requires MFA, logging)
Monitoring (detect anomalous token issuance patterns)

Token Issuance Monitoring: Authorization Server SHOULD monitor and alert on: - Sudden increase in token issuance rate (possible compromise) - Tokens with unusual capability combinations (possible policy bypass) - Tokens issued to unknown agents (possible credential theft) - Issuance outside normal business hours (possible unauthorized access)

Resource Server Security Input Validation: Resource Servers MUST validate: - Token signature with trusted AS public key - Token expiration (exp) with acceptable clock skew (RECOMMENDED: ≤5 minutes) - Audience (aud) matches Resource Server identifier - Agent identity (agent.id) is recognized or allowed by policy - All constraints in matching capability Error Handling (Privacy-Preserving): Resource Servers MUST NOT leak authorization details in error responses: Bad (leaks constraint values): json { "error": "aap_constraint_violation", "error_description": "Rate limit exceeded: 51 requests when max is 50" } Good (privacy-preserving): json { "error": "aap_constraint_violation", "error_description": "Request violates capability constraints" } Detailed violation information SHOULD be logged server-side for audit, not returned to client.

Additional Security Considerations Token Logging:

Tokens MUST NOT be logged in plaintext in application logs
Use token redaction in logging libraries (replace with hash or truncated value)
If logging is necessary for debugging, use separate secure audit log with strict access control

Clock Skew:

Resource Servers SHOULD tolerate up to 5 minutes of clock skew for exp and nbf validation
Skew tolerance MUST NOT exceed 5 minutes (to limit window for expired token use)
Organizations SHOULD use NTP or equivalent for clock synchronization

Compliance Considerations: AAP enables compliance with regulations requiring: - Explicit authorization (GDPR Article 6) - Purpose limitation (GDPR Article 5) - Audit trails (SOC 2, ISO 27001, HIPAA) - Access controls (PCI-DSS, FedRAMP) However, AAP alone is not sufficient for compliance; organizational policies, procedures, and controls are also required.

Privacy Considerations AAP tokens and audit logs may contain information that could identify agents, organizations, individuals, or sensitive operational details. This section provides guidance on privacy protection in AAP implementations.

Personal Data in AAP Tokens AAP tokens MAY contain personal data under GDPR, CCPA, and similar privacy regulations: Potentially Personal Claims: - agent.operator: Organization identifier (may be personal if sole proprietor or individual developer) - task.created_by: User ID who initiated the task - task.purpose: May contain user-identifying details if not carefully crafted - audit.trace_id: Potentially correlatable across requests (tracking) - delegation.chain: Reveals agent interaction patterns and organizational structure - context.location: Geographic location information Data Controller and Processor: - Authorization Server operator is typically the data controller for token claims - Resource Server operators are data processors when validating tokens - Delegation across organizations may create complex controller/processor relationships

Data Minimization Principles Implementations SHOULD apply the principle of data minimization [GDPR Article 5(1)(c)]: Guideline: Include Only What Is Necessary

Claim	Privacy-Preserving Approach	Privacy-Risky Approach
`agent.id`	Use pseudonymous ID: `urn:uuid:550e8400-...`	Use descriptive name: `agent-for-alice@example.com`
`task.id`	Use UUID: `550e8400-e29b-41d4-...`	Use descriptive ID: `task-research-for-alice-project-secret`
`task.purpose`	Use category: `research` or `content-creation`	Use full description: `Research climate data for Alice's PhD thesis on...`
`task.created_by`	Use pseudonymous ID: `user:u123456`	Use email: `alice.smith@example.com`
`audit.trace_id`	Use per-task UUID (rotated)	Use stable agent-wide ID (enables long-term correlation)

Unnecessary Claims SHOULD Be Omitted: - Don't include agent.name if agent.id is sufficient for authorization - Don't include task.metadata unless necessary for constraint enforcement - Don't include context.location.ip_address unless geofencing is required

Retention and Lifecycle Token Retention: - Tokens expire per exp claim (typically minutes to hours) - Expired tokens SHOULD be immediately discarded by agents - Tokens SHOULD NOT be persisted to disk (memory-only storage) Audit Log Retention: - Audit logs SHOULD follow organizational retention policy - Minimum retention: As required by compliance framework (e.g., SOC 2: 1 year) - Maximum retention: Only as long as necessary for audit and investigation - After retention period: Logs MUST be securely deleted or anonymized Delegation Chain Retention: - Delegation chains SHOULD NOT be retained in logs beyond audit window - When logging delegation events, consider hashing agent IDs rather than storing plaintext Right to Erasure (GDPR Article 17): - If task.created_by contains personal data, organizations MUST support erasure requests - Consider using pseudonymous IDs that can be de-linked from personal data - Audit logs may be exempt from erasure if required for legal compliance; consult legal counsel

Cross-Domain Correlation Threat: Malicious Resource Servers in different organizations could correlate requests across services using stable identifiers in AAP tokens (e.g., audit.trace_id, agent.id). Impact: - Agent behavior profiling (what capabilities, what resources, what patterns) - Competitive intelligence (infer agent strategies from request patterns) - Privacy violation (correlate agent activity across unrelated services) Mitigations: Trace ID Rotation (REQUIRED for Cross-Domain): - When token audience changes trust domain, generate new audit.trace_id - Example: Token for api.example.com has trace_id: abc123; delegated token for external-service.org has trace_id: xyz789 - Correlation within single organization preserved; cross-organization correlation prevented Domain-Specific Agent IDs: - Use different agent.id per Resource Server trust domain - Example: Agent presents agent:internal-001 to internal APIs, agent:partner-facing-alpha to partner APIs - Authorization Server maintains mapping; Resource Servers see domain-specific IDs Minimize Identifiable Information in Delegated Tokens: - When delegating to external tool, remove non-essential claims - Example: Strip task.created_by, context.location when crossing trust boundary - Token Exchange request SHOULD specify required_claims (minimal set) Trace ID Scope Claim (RECOMMENDED): json "audit": { "trace_id": "550e8400-e29b-41d4-a716-446655440000", "trace_id_scope": "task" // or "session", "agent", "domain" } This signals the intended correlation boundary: - task: Trace ID unique per task (rotated between tasks) - session: Trace ID unique per agent session (rotated on agent restart) - agent: Trace ID stable for agent lifetime (enables long-term correlation; privacy-risky) - domain: Trace ID unique per trust domain (rotated when crossing domains)

Privacy-Preserving Error Messages Resource Servers MUST NOT leak authorization details in error responses that could enable privacy violations or capability profiling. Avoid: - "Agent does not have capability 'delete.data'" (reveals capabilities) - "Rate limit: 51 requests, max allowed 50" (reveals constraint values) - "Domain blocked: malicious.com is in blocklist" (reveals policy details) - "Task purpose 'research for Alice' does not match action 'publish'" (reveals task details) Prefer: - "Insufficient permissions" (generic, privacy-preserving) - "Request violates capability constraints" (doesn't specify which constraint) - "Authorization failed" (minimal information disclosure) Detailed Error Information: - Log server-side with full details (for audit and debugging) - Include error correlation ID in response for support tickets - Client can reference correlation ID when contacting support; support team accesses server logs Example Privacy-Preserving Error Response: json { "error": "insufficient_permissions", "error_description": "The request could not be authorized", "error_correlation_id": "err-550e8400-e29b-41d4-a716-446655440000" } Server log (not returned to client): json { "error_correlation_id": "err-550e8400-e29b-41d4-a716-446655440000", "error": "aap_constraint_violation", "constraint_violated": "domains_allowed", "requested_domain": "malicious.com", "allowed_domains": ["example.org"], "agent_id": "agent-researcher-01", "task_id": "task-12345", "timestamp": "2024-01-01T12:00:00Z" }

Anonymization and Pseudonymization Pseudonymization (GDPR Article 4(5)): Pseudonymization is the processing of personal data such that it can no longer be attributed to a specific data subject without additional information (kept separately and secured). AAP Pseudonymization Techniques:

Data Type	Pseudonymization Approach
User IDs (`task.created_by`)	Use UUID mapped to real user ID in separate database
Agent names	Use `agent:0001` instead of `agent-for-alice`
Task purposes	Use task category codes instead of free-text descriptions
Trace IDs	Use cryptographic hash of (user ID + task ID + salt)

Anonymization (Irreversible): For audit logs past retention period: - Replace agent.id with hash (if agent identity no longer needed) - Remove task.created_by entirely - Aggregate statistics (e.g., "1000 requests by research agents") instead of individual records

Consent and Transparency Agent Operator Transparency: When an agent acts on behalf of a human user: - User SHOULD be informed that an agent will perform actions - User SHOULD be shown the task purpose and capabilities granted - User SHOULD be able to review and revoke agent authorizations Example User Notification: ``` Your request to "research climate change impacts" has been assigned to an AI agent. The agent will be able to: - Search web resources from example.org (max 50 requests/hour) - Create draft articles in the CMS The agent will NOT be able to: - Publish articles (requires your approval) - Access data outside example.org - Perform actions unrelated to research You can revoke this authorization at any time in your account settings. ``` Token Transparency: For compliance with transparency requirements (GDPR Article 13-14): - Organizations SHOULD document what claims are included in AAP tokens - Organizations SHOULD inform users when their actions trigger agent authorization - Organizations SHOULD provide access to audit logs (subject to security and legal constraints)

Third-Party Tool Privacy When delegating to third-party tools (external organizations): Data Sharing Agreement: - Delegation to external tool constitutes data sharing - Organizations SHOULD have data processing agreements (DPAs) with tool providers - Token Exchange to external AS SHOULD trigger data sharing notification/consent Claim Filtering: - Authorization Server SHOULD filter claims when delegating to external tool - Remove non-essential claims: task.created_by, agent.operator internal details - Retain only authorization-essential claims: capabilities, delegation.depth Example Filtered Delegation: Original token (internal): json { "agent": {"id": "agent-001", "operator": "org:acme-corp"}, "task": {"id": "task-123", "purpose": "research", "created_by": "user:alice"}, "capabilities": [{"action": "search.web"}] } Delegated token (external tool): json { "agent": {"id": "delegated-from:acme-corp", "operator": "org:acme-corp"}, "task": {"id": "task-123-external", "purpose": "research"}, // "created_by" removed for privacy "capabilities": [{"action": "search.web", "constraints": {"domains_allowed": ["example.org"]}}], "delegation": {"depth": 1, "chain": ["agent-001", "external-tool"]} }

Privacy by Design Recommendations For Authorization Server Implementations:

Default to short trace ID rotation (per-task, not per-agent)
Provide configuration options for privacy levels (minimal, standard, full disclosure)
Support claim filtering on Token Exchange
Log privacy-impacting events (cross-domain delegation, long trace ID retention)

For Resource Server Implementations:

Never log tokens in plaintext
Redact personal data in error responses
Provide audit log access controls (only authorized personnel)
Support audit log anonymization after retention period

For Agent Implementations:

Request minimal capabilities (least privilege reduces privacy risk)
Specify minimal required_claims on Token Exchange
Rotate trace IDs when crossing trust boundaries
Discard tokens immediately upon expiration (don't cache expired tokens)

Regulatory Compliance Guidance GDPR Compliance: - AAP supports GDPR principles: purpose limitation (task binding), data minimization (constrained capabilities), accountability (audit logs) - Organizations MUST implement additional controls: consent management, data subject rights, DPIAs CCPA Compliance: - AAP audit logs may constitute "personal information" if they identify consumers - Organizations MUST support consumer rights: access, deletion, opt-out of sale - Consider using pseudonymous IDs to reduce CCPA applicability HIPAA Compliance (Healthcare): - AAP tokens accessing Protected Health Information (PHI) MUST use proof-of-possession - Audit logs MUST meet HIPAA retention requirements (6 years) - Delegation chains provide required access audit trail Note: AAP is an authorization protocol, not a complete privacy framework. Organizations MUST implement privacy policies, procedures, and controls beyond AAP technical mechanisms.

IANA Considerations This document has no IANA actions. If in the future the AAP claim names (aap_agent, aap_task, aap_capabilities, aap_oversight, aap_delegation, aap_context, aap_audit) are registered in the "JSON Web Token Claims Registry" as defined in , the required registration procedure would be followed.

Related Work and Comparison This section positions AAP within the broader ecosystem of authorization, access control, and identity systems. AAP is designed to complement and extend existing standards rather than replace them.

OAuth 2.0 Scopes Traditional OAuth 2.0 scopes provide coarse-grained authorization through simple string tokens (e.g., read:articles, write:cms). While effective for user-delegated access, scopes have limitations for autonomous agent scenarios: OAuth Scopes Characteristics: - Simple string-based permissions - No built-in constraint mechanism - No task context binding - No delegation tracking - Limited expressiveness for complex policies AAP Enhancements: - Structured capabilities: Actions (search.web) with typed, enforceable constraints (max_requests_per_hour: 50, domains_allowed: ["example.org"]) - Task binding: Scopes don't express purpose or context; AAP tokens bind authorization to specific tasks with declared purposes - Delegation chains: OAuth has no built-in delegation tracking; AAP provides delegation.depth and delegation.chain for auditable multi-hop authorization - Operational limits: Scopes are binary (granted/not granted); AAP constraints enable quantitative limits (rate, volume, time windows) Comparison Example:

Scenario	OAuth Scope Approach	AAP Approach
Research agent web scraping	`scope=read:web` (grants unlimited web access)	`capabilities: [{"action": "search.web", "constraints": {"domains_allowed": ["example.org"], "max_requests_per_hour": 50}}]`
Content creation with approval	`scope=write:cms` (cannot distinguish draft vs. publish)	`capabilities: [{"action": "cms.create_draft"}]` with `oversight: {"requires_human_approval_for": ["cms.publish"]}`
Delegated tool calls	Requires new OAuth flow per delegation	Token Exchange with automatic `delegation.depth` increment and privilege reduction

Compatibility: AAP maintains backward compatibility by optionally including OAuth scope claim alongside AAP capabilities.

Fine-Grained Authorization Systems (Zanzibar, ReBAC) Google Zanzibar and Relation-Based Access Control (ReBAC) systems focus on resource-level permissions with global consistency ("user X can read document Y based on relationship graph"). Zanzibar/ReBAC Characteristics: - Resource-centric authorization - Graph-based relationship evaluation - Highly scalable, globally consistent - Focus: "Who can access what resource?" AAP Focus: - Agent-centric authorization - Task context and operational constraints - Delegation chain tracking - Focus: "What can this agent do, under what conditions, for what purpose?" Complementary Use: AAP and Zanzibar address different layers: - AAP: Authorizes the agent to call APIs, binds actions to tasks, enforces operational limits - Zanzibar: Determines whether the agent (once authorized) can access specific resources Integration Pattern: 1. Agent presents AAP token to API (capability: "document.read") 2. API validates AAP token (RS validation) 3. API calls Zanzibar to check if agent can read specific document 4. Both checks must pass for request to succeed

Cloud Identity and Access Management Major cloud providers offer identity and temporary credential systems: AWS Security Token Service (STS): - Cross-account delegation with session policies - Temporary credentials with limited lifetime - AssumeRole for privilege escalation/reduction - IAM policies define permissions GCP Workload Identity: - Kubernetes ServiceAccount federation to GCP IAM - Workload identity binding for containerized applications - OAuth 2.0-based token exchange AAP Differences: - Vendor-neutral: AAP is not tied to a specific cloud provider - Agent-specific claims: agent.model, task.purpose, oversight requirements not present in cloud IAM - Explicit delegation tracking: Cloud IAM tracks roles/sessions but not multi-hop agent delegation chains - Task binding: Cloud credentials are session-bound, not task-bound Integration: AAP can be used as a policy enforcement layer above cloud IAM. For example: - AAP token authorizes agent for a task - Agent exchanges AAP token for cloud-specific credentials (AWS STS, GCP token) - Cloud IAM enforces resource-level permissions

Service Mesh Authorization (Istio, Linkerd) Service meshes provide network-level security with mutual TLS and Layer 7 policies: Service Mesh Capabilities: - mTLS for service-to-service encryption - SPIFFE/SPIRE for workload identity - Authorization policies based on service identity - Network-level request routing and policy enforcement AAP Capabilities: - Application-level agent identity with model and operator metadata - Task semantics and purpose binding - Business logic constraints (rate limits, domain allowlists, approval workflows) - Delegation chain for multi-agent workflows Complementary Layers:

Layer	Technology	What It Provides
Network (L4/L7)	Service Mesh	Identity (SPIFFE), mTLS, service-to-service authz
Application	AAP	Agent identity, task binding, capability constraints, oversight

Integration: - Use SPIFFE IDs in AAP agent.id claim - Service mesh enforces network policy ("can agent reach API?") - AAP enforces business logic ("can agent perform this action with these constraints?")

Capability-Based Security AAP uses the term "capability" from classic capability-based security literature (Dennis & Van Horn 1966, Levy 1984): Classic Capability Systems: - Unforgeable tokens (capabilities) that grant access to specific objects - Capabilities combine authority and designation - No ambient authority (no ACLs checked at access time) - Pure object-capability model (E language, seL4 microkernel) AAP Capabilities: - Structured claims in signed JWTs representing permissions - Combine action designation (search.web) with constraints - Rely on centralized Authorization Server (not pure capability model) - Inherit unforgeability from JWT signature (not object references) Key Difference: AAP is not a pure object-capability system. It uses centralized issuance (AS) and validation (RS) rather than distributed capability passing. However, AAP borrows the principle of least authority: each capability explicitly states what is allowed and under what constraints, rather than checking ambient permissions.

OpenID Connect and Step-Up Authentication OpenID Connect (OIDC) provides identity claims and authentication strength signaling: OIDC Capabilities: - User identity claims (sub, email, name) - Authentication context (acr, amr claims) - Step-up authentication via acr_values parameter - ID tokens separate from access tokens AAP Capabilities: - Agent identity (non-human, autonomous software) - Task binding and operational constraints - Oversight requirements signal when human approval needed - Access tokens include both identity and authorization Overlap: Both use JWT and OAuth 2.0 foundation. Integration: - AAP oversight.requires_human_approval_for can trigger OIDC step-up authentication - Resource Server can request user approval via OIDC interaction, using acr_values for higher assurance - Human supervisor identified in oversight.supervisor can authenticate via OIDC

OAuth 2.0 Rich Authorization Requests (RAR) Rich Authorization Requests [draft-ietf-oauth-rar] extend OAuth to support complex authorization requirements beyond simple scopes. RAR Features: - Structured authorization details in JSON - Request-specific authorization (e.g., payment with amount) - Detailed, fine-grained permissions AAP vs. RAR: - RAR: Request-time authorization details sent by client - AAP: Token-time structured claims issued by AS, enforced by RS - RAR: Focuses on authorization request - AAP: Focuses on token structure and enforcement Complementary: RAR can be used to request AAP capabilities. Client sends RAR with desired capabilities; AS issues AAP token with granted capabilities.

Summary: Where AAP Fits AAP occupies a unique position in the authorization ecosystem:

System	Focus	AAP Relationship
OAuth Scopes	Coarse-grained API access	AAP extends with structured capabilities
Zanzibar/ReBAC	Resource-level permissions	AAP handles agent/task layer; Zanzibar handles resource layer
Cloud IAM	Cloud resource access	AAP is vendor-neutral layer; can integrate with cloud IAM
Service Mesh	Network-level security	AAP adds application-level semantics (task, constraints)
OIDC	User identity	AAP focuses on agent identity; integrates for human oversight
RAR	Authorization requests	AAP defines token structure; RAR can request AAP capabilities

AAP is designed for scenarios where: - The client is an autonomous AI agent (not a human user) - Actions must be bound to explicit tasks with declared purposes - Authorization requires operational constraints (rate limits, domain restrictions, time windows) - Delegation chains must be tracked and auditable - Human oversight is required for high-risk actions - Standard OAuth scopes are insufficient for expressing agent policies

Field	Original Token	Derived Token	Change Type
`aud`	`https://api.example.com`	`https://tool-scraper.example.com`	Audience changed
`exp`	1704067200 (3600s)	1704065400 (1800s)	Lifetime reduced 50%
`jti`	`token-original-123`	`token-delegated-456`	New unique ID
`capabilities`	2 capabilities	1 capability	`cms.create_draft` removed
`capabilities[0].constraints.domains_allowed`	`["example.org", "trusted.com"]`	`["example.org"]`	Domain list narrowed
`capabilities[0].constraints.max_requests_per_hour`	100	50	Rate limit reduced
`delegation.depth`	0	1	Depth incremented
`delegation.chain`	1 entry	2 entries	Tool appended to chain
`delegation.parent_jti`	(absent)	`token-original-123`	Parent link added

Error Code	HTTP Status	Description	Example Scenario
`aap_invalid_capability`	403 Forbidden	No matching capability for requested action	Agent has `search.web` but requests `cms.publish`
`aap_constraint_violation`	429 Too Many Requests or 403 Forbidden	Capability constraint violated	51st request when `max_requests_per_hour: 50`
`aap_task_mismatch`	403 Forbidden	Request inconsistent with task purpose	Token for "research" task used for data deletion
`aap_approval_required`	403 Forbidden	Action requires human approval	`cms.publish` in `oversight.requires_human_approval_for`
`aap_excessive_delegation`	403 Forbidden	Delegation depth exceeded	Token with `depth=3` when `max_depth=2`
`aap_invalid_context`	403 Forbidden	Context restriction violated	Request outside `time_window` or from blocked region
`aap_domain_not_allowed`	403 Forbidden	Target domain not in allowlist	Request to `malicious.com` when `domains_allowed: ["example.org"]`
`aap_agent_not_recognized`	403 Forbidden	Agent identity not recognized by policy	Unknown `agent.id` value
`aap_invalid_delegation_chain`	403 Forbidden	Delegation chain validation failed	Chain length doesn't match depth, or parent token invalid
`aap_capability_expired`	403 Forbidden	Time-based capability expired	Request after `time_window.end`