Agent Identity Protocol: Agentic Authentication and Authorized Policy Enforcement

3.1. The Identity Gap

When an AI agent calls a tool, it presents credentials that belong to a human account. The tool server cannot tell whether the actor is a human or an agent, which agent it is, or what limits apply to it. This creates four concrete problems:¶

(a)

Security -- A compromised or manipulated agent can invoke any¶

tool the human account can reach. There is no agent-specific authorization boundary.¶

(b)

Auditability -- Logs record actions against a human account.¶

After an incident, investigators cannot determine which actions were taken by a human versus an agent.¶

(c)

Compliance -- Regulations increasingly require traceability of¶

automated decision-making. Without agent-level identity, organizations cannot satisfy these requirements.¶

(d)

Accountability -- Billing, rate limits, and quotas are scoped¶

to human accounts. Agent usage cannot be isolated or attributed.¶

3.2. The Enforcement Gap

Even where agent behavior policies exist, they are expressed as text in model system prompts. System prompts are not tamper-evident and can be bypassed by adversarial inputs to the model. There is no infrastructure-layer enforcement point that acts independently of the model.¶

3.3. Scope of This Specification

AIP closes both gaps. Layer 1 gives every agent a distinct, verifiable identity independent of the human principal's credentials. Layer 2 enforces agent-specific policy at the tool-call boundary, outside the model's trust domain, in a way that cannot be overridden by model outputs.¶

4.1. Architecture

AIP introduces two components into the agent-to-tool-server path: an AIP Registry (Layer 1) and an AIP Proxy (Layer 2).¶

                 +-----------------------------------------+
                 |           AI Client (Agent)             |
                 +-----------------+-----------------------+
                                   | tools/call + AIP Token
                                   v
                 +-----------------------------------------+
                 |            AIP Proxy (Layer 2)          |
                 |  +-----------+  +--------------------+  |
                 |  |  Verify   |  |   Policy Engine    |  |
                 |  |  Token    |  |   (AgentPolicy)    |  |
                 |  +-----+-----+  +---------+----------+  |
                 |        |                  |              |
                 |  +-----v------------------v----------+  |
                 |  |         Audit Logger              |  |
                 |  +-----------------------------------+  |
                 +-----+-------------------+---------------+
                       |                   |
                ALLOW  |             DENY  |  HOLD
                       v                   v
                 +-----------+    +------------------+
                 | Tool      |    | Error / HITL     |
                 | Server    |    | Queue            |
                 +-----------+    +------------------+

                 ^
                 | key lookup
                 |
            +---------+
            |   AIP   |
            | Registry|
            | (Lay. 1)|
            +---------+

4.2. Call Lifecycle

The lifecycle of a single tool call under AIP is as follows:¶

1. Registration (once, at deploy time). The principal registers the agent with an AIP Registry. The registry assigns an Agent ID and stores the agent's public key. The agent stores its private key securely.¶

2. Token construction (per call). Before each tool call the agent constructs an AIP Token (Section 5.6): a small JSON object containing the Agent ID, tool name, argument hash, nonce, timestamp, and a key signature over the token.¶

3. Proxy intercept. The AIP Proxy receives the tool call request before it reaches the tool server.¶

4. Token verification (Layer 1). The proxy retrieves the agent's public key from the registry (or local cache) and verifies the signature. It also checks the nonce for replay and the timestamp for freshness.¶

5. Policy evaluation (Layer 2). The proxy checks the call against the AgentPolicy: is the tool on the allowlist? Do the arguments pass validation? Does the call trigger a HITL hold? Does the response contain sensitive data that must be redacted?¶

6. Decision. The proxy produces one of three outcomes: ALLOW -- forward the call to the tool server; DENY -- return an error to the agent, tool server not reached; HOLD -- queue the call for human approval.¶

7. Audit. Every decision is written to the audit log regardless of outcome.¶

4.3. Relationship to Existing Standards

• OAuth 2.0 [RFC6749] / OIDC [OIDC]: AIP does not replace service- level authentication. The agent still presents its OAuth token or API key to the tool server. AIP provides a separate agent identity layer that the proxy can verify independently.¶

• SPIFFE/SVID [SPIFFE]: In Kubernetes deployments, the AIP Proxy MAY use a SPIFFE SVID to authenticate to tool servers over mTLS, layering workload identity on top of AIP agent identity.¶

• JSON-RPC 2.0 [JSON-RPC]: Tool call and error messages use the JSON-RPC 2.0 wire format, compatible with MCP and similar protocols.¶

5.1. Agent Registration

An agent is provisioned by its principal submitting a registration request to an AIP Registry. The request MUST include:¶

(a)

The agent's public key, base64url-encoded [RFC4648];¶

(b) The principal identifier -- a string that uniquely identifies the accountable human or organization (e.g., an email address, an organization slug, or an OAuth subject claim); (TO BE WORKED ON FURTHER)¶

(c)

A human-readable agent name (RECOMMENDED);¶

(d) An optional free-text description.¶

The registration request MUST be authenticated. Authentication MAY be via OAuth 2.0 bearer token, mTLS client certificate, or a pre- shared registration secret, at the registry operator's discretion.¶

On successful registration, the registry:¶

(a)

Assigns a UUID v4 [RFC4122] as the agent's local identifier;¶

(b) Constructs the Agent ID as "<registry-host>/<uuid>"; (c) Stores the Agent Record (Section 5.2); (d) Returns the Agent ID to the principal.¶

The principal MUST store the Agent ID and configure the agent with both the Agent ID and its private key before deployment.¶

5.2. Agent Record

The Agent Record is the data structure the AIP Registry stores for each registered agent. It MUST contain:¶

agentId (string):

The Agent ID assigned at registration, in the form "<registry-host>/<uuid-v4>". Example: "reg.agentidentityprotocol.io/01933f4a-9b2c-7d8e-af01"¶

publicKey (string):

The agent's current public key, base64url-encoded.¶

principalId (string):

Identifier of the accountable principal.¶

name (string):

Human-readable agent name. Informational only; not authenticated.¶

description (string, optional):

Free-text description of the agent's purpose.¶

createdAt (string):

ISO 8601 UTC timestamp of registration.¶

keyHistory (array):

Append-only list of all public keys ever bound to this Agent ID. Each entry contains "publicKey", "activeFrom", and "revokedAt" (null if still active).¶

status (string):

One of "active" or "revoked".¶

Example Agent Record (JSON):¶

{
  "agentId": "reg.agentidentityprotocol.io/01933f4a-9b2c-7d8e-af01",
  "publicKey": "MCowBQYDK2VwAyEAz8vG...",
  "principalId": "acme-corp",
  "name": "ResearchAssistant-v1",
  "description": "Internal document retrieval agent",
  "createdAt": "2026-01-15T09:00:00Z",
  "keyHistory": [
    {
      "publicKey": "MCowBQYDK2VwAyEAz8vG...",
      "activeFrom": "2026-01-15T09:00:00Z",
      "revokedAt": null
    }
  ],
  "status": "active"
}

5.3. AIP Registry API

The AIP Registry MUST expose the following HTTP endpoints over TLS 1.3 (Section 9.4):¶

POST   /v1/agents              Register a new agent
GET    /v1/agents/{agentId}    Retrieve an Agent Record
PUT    /v1/agents/{agentId}/key  Rotate the agent's public key
DELETE /v1/agents/{agentId}    Revoke an agent
GET    /v1/revocations/stream  SSE stream for revocation events

The GET /v1/agents/{agentId} endpoint is the only endpoint that MUST be reachable by AIP Proxies at call-verification time. All other endpoints are used during provisioning and key management.¶

Responses from GET /v1/agents/{agentId} MUST include the full Agent Record. Proxies SHOULD cache this response for at least 30 seconds. The cache MUST be invalidated on receipt of a revocation event from the SSE stream.¶

5.4. Key Rotation

A principal rotates an agent's key by submitting a PUT request to /v1/agents/{agentId}/key, authenticated with the current private key. The request body MUST contain the new public key.¶

The registry MUST:

(a) Set "revokedAt" on the current keyHistory entry to the current timestamp;¶

(b)

Append a new keyHistory entry with the new public key;¶

(c) Update the "publicKey" field on the Agent Record; (d) Emit a rotation event on the revocations SSE stream so that proxies can invalidate their cached Agent Records immediately.¶

The Agent ID does NOT change on key rotation.¶

5.5. Agent Revocation

A principal revokes an agent by sending a DELETE request to /v1/agents/{agentId}. The registry MUST set the agent's status to "revoked" and emit a revocation event on the SSE stream.¶

Proxies MUST reject AIP Tokens from revoked agents with error AIP-E012. Revoked Agent Records MUST be retained in the registry for audit log verification but MUST NOT be returned as "active".¶

5.6. The AIP Token (Agent Attestation Token)

The AIP Token is a compact signed JSON object that the agent constructs and attaches to every outbound tool call. It is the mechanism by which the agent asserts its identity to the proxy.¶

{
  "aipVersion": "1",
  "agentId": "reg.agentidentityprotocol.io/01933f4a-9b2c-7d8e-af01",
  "tool": "read_file",
  "argumentsHash": "e3b0c44298fc1c149afb4c8996fb924...",
  "nonce": "a3f8b2c1d4e5f607a8b9c0d1e2f3a4b5",
  "timestamp": "2026-02-24T14:30:00Z",
  "signature": "TUlJQ0lqQU5CZ2txaGtpRzl3MEJB..."
}

5.7. Token Verification

The AIP Proxy MUST perform the following checks in order. A failure at any step MUST produce a DENY decision with the corresponding error code (Section 7.2) and an audit log entry.¶

Step 1 -- Presence check.

Confirm the AIP Token is present in the request. On failure: AIP-E010.¶

Step 2 -- Agent Record lookup.

Resolve the Agent ID from the token against the AIP Registry (or local cache). Confirm the Agent Record status is "active". On failure: AIP-E011 (unresolvable) or AIP-E012 (revoked).¶

Step 3 -- Signature verification.

Compute the canonical serialization (Section 5.6.2) and verify the key signature against the public key in the Agent Record. This operation MUST be performed in constant time. On failure: AIP-E013.¶

Step 4 -- Nonce replay check.

Check the nonce against the proxy's local nonce cache (minimum TTL: 600 seconds). If the nonce has been seen before, reject. On failure: AIP-E004.¶

Step 5 -- Timestamp freshness.

Reject the token if its timestamp is more than 300 seconds in the past or 30 seconds in the future relative to the proxy clock. On failure: AIP-E005.¶

If all five steps pass, the token is verified and the call proceeds to policy evaluation (Section 6.3).¶

6.1. Proxy Architecture

The AIP Proxy is a forward proxy interposed between the AI client and one or more tool servers. It is transparent: it presents itself to the AI client as the tool server endpoint, and to the tool server as an authorized caller.¶

The proxy operates in one of two modes, set per agent in the AgentPolicy:¶

enforce:

Policy violations result in DENY responses. The tool¶

server is never reached for blocked calls. This is the default and RECOMMENDED mode.¶

monitor:

Policy violations are logged but calls are forwarded¶

regardless. Used for baselining and policy development before switching to enforce mode.¶

The proxy MUST maintain locally:¶

o  The AgentPolicy for each agent it serves;
o  A bounded nonce cache (TTL >= 600 seconds, LRU eviction);
o  A revocation cache (refresh interval <= 60 seconds);
o  An append-only audit log.

6.2. AgentPolicy

The AgentPolicy is a YAML file that declares the enforcement rules for a specific agent. One AgentPolicy file corresponds to one Agent ID. A AgentPolicy file can apply to multiple Agent IDs.¶

<CODE BEGINS>
   agentId: <Agent ID>
   mode: <enforce | monitor>

   tools:
     allowed:
       - <tool-name>
     rules:
       - tool: <tool-name>
         action: <allow | ask | block>
         args:
           <arg-name>:
             pattern: <PCRE regex>
             maxLength: <integer>

   dlp:
     - name: <pattern name>
       regex: <PCRE regex>
       action: <redact | block>
       scope: <request | response | both>

   hitl:
     approvers:
       - <email or identifier of approver>
     timeout_seconds: <integer, default 300>  # seconds to wait for approval
     on_timeout: <deny | allow>               # action if no response received

<CODE ENDS>

6.3. Intercept Flow

The following is the normative sequence for every tool call received by the proxy.¶

1.  Receive the tool call from the AI client.

2.  Run token verification (Section 5.7).
    Any failure -> DENY with the corresponding AIP-Exxx code.

3.  Check tools.allowed.
    Tool not present and mode == enforce -> DENY AIP-E001.

4.  Check tools.rules for this tool.
    action == block -> DENY AIP-E003.
    action == ask   -> go to step 7 (HITL).

5.  Validate arguments against any matching tools.rules.args.
    Violation -> DENY AIP-E002.

6.  Run DLP scan on request arguments.
    block match -> DENY AIP-E008.
    redact match -> replace content, continue.

7.  If action == ask: submit to HITL queue (Section 6.5).
    Approved -> continue to step 8.
    Denied   -> DENY AIP-E015.
    Timed out -> resolve per on_timeout.

8.  ALLOW: forward the call to the tool server.

9.  Receive the tool server response.

10. Run DLP scan on the response.
    block match -> return AIP-E008 to agent, suppress response.
    redact match -> replace content, continue.

11. Return the (possibly redacted) response to the agent.¶

12. Write audit log record (Section 6.7).¶

6.4. Decision Outcomes

ALLOW:

The call passed all checks and was forwarded. The tool¶

server response was returned to the agent.¶

DENY:

The call was rejected before reaching the tool server.¶

The proxy returns a JSON-RPC 2.0 error (Section 7.2). The tool server receives nothing.¶

HOLD:

The call is queued. The proxy returns a pending response¶

to the agent. The call is resolved when a human approver responds or the HITL timeout is reached.¶

6.5. Human-in-the-Loop (HITL)

When a call is held, the proxy MUST notify all addresses listed in hitl.approvers. The notification MUST include:¶

• The Agent ID and agent name;¶

• The tool name and (post-redaction) arguments;¶

• The policy rule that triggered the hold;¶

• A unique hold_id.¶

Notification delivery (email, webhook, Slack, web UI) is out of scope for this specification.¶

Approvers submit a decision to the proxy using the HITL API. Approval and denial endpoints are:¶

   POST /v1/hitl/{hold_id}/approve
   POST /v1/hitl/{hold_id}/deny

These endpoints MUST be authenticated. If no response is received within hitl.timeout_seconds, the proxy resolves the hold according to hitl.on_timeout and writes the outcome to the audit log.¶

6.6. Data Loss Prevention (DLP)

The DLP scanner applies the dlp rules from the AgentPolicy to both the inbound tool call arguments and the outbound tool server response. Rules are evaluated in the order they are listed. The first matching rule wins.¶

Implementations SHOULD provide a standard rule library covering common sensitive data types:¶

• Cloud provider credentials (AWS, GCP, Azure key patterns);¶

• Private key material (PEM headers);¶

• Common PII patterns (email, phone, SSN formats);¶

• Generic high-entropy secrets (long alphanumeric tokens).¶

6.7. Audit Logging

The proxy MUST write one log record per tool call outcome. Records MUST be appended to an append-only log and MUST NOT be modified after writing.¶

Log records SHOULD be hash-chained: each record includes the SHA-256 hash of the previous record, enabling tamper detection.¶

See Section 7.3 for the normative record format.¶

Log records MUST be retained for a minimum of 90 days.¶

7.1. AIP-Token Header

For HTTP-based tool transports, the AIP Token is conveyed in the request header:¶

AIP-Token: <base64url-encoded UTF-8 JSON token>¶

base64url encoding is defined in [RFC4648] Section 5 (no padding).¶

For MCP stdio transports, the AIP Token is conveyed as a "_aip" field at the top level of the JSON-RPC request object:¶

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "id": 1,
  "params": {
    "name": "read_file",
    "arguments": { "path": "/data/report.txt" }
  },
  "_aip": {
    "aipVersion": "1",
    "agentId": "reg.agentidentityprotocol.io/01933f4a...",
    "tool": "read_file",
    "argumentsHash": "e3b0c44298...",
    "nonce": "a3f8b2c1d4e5f607...",
    "timestamp": "2026-02-24T14:30:00Z",
    "signature": "TUlJQ0lq..."
  }
}

7.2. Error Response Format

Error responses MUST conform to JSON-RPC 2.0 [JSON-RPC]. AIP error codes are in the range -32001 to -32099:¶

Code     ID        Meaning
------   --------  -----------------------------------------------
-32001   AIP-E001  Tool not in allowlist
-32002   AIP-E002  Argument validation failed
-32003   AIP-E003  Tool unconditionally blocked
-32004   AIP-E004  Nonce replay detected
-32005   AIP-E005  Timestamp out of range
-32008   AIP-E008  DLP violation
-32010   AIP-E010  AIP Token missing
-32011   AIP-E011  Agent ID not found in registry
-32012   AIP-E012  Agent revoked
-32013   AIP-E013  Signature verification failed
-32015   AIP-E015  HITL approval denied
-32016   AIP-E016  HITL timed out
-32099   AIP-E099  Internal proxy error

Example error response:

{
  "jsonrpc": "2.0",
  "id": 1,
  "error": {
    "code": -32001,
    "message": "AIP-E001: tool not in allowlist",
    "data": {
      "aipCode": "AIP-E001",
      "agentId": "reg.agentidentityprotocol.io/01933f4a...",
      "tool": "exec_command"
    }
  }
}

7.3. Audit Log Record

Each record is a single JSON object on one line (JSONL [JSONL]):¶

{
  "v": 1,
  "ts": "<ISO 8601 UTC>",
  "eventId": "<UUID v4>",
  "prevHash": "<SHA-256 hex of previous record, or null>",
  "decision": "<ALLOW | DENY | HOLD>",
  "errorCode": "<AIP-Exxx or null>",
  "agentId": "<Agent ID>",
  "principalId": "<principal identifier>",
  "tool": "<tool name>",
  "argumentsHash": "<SHA-256 hex>",
  "policyName": "<AgentPolicy agentId value>",
  "verificationStep": "<1-5 or null if passed>",
  "dlp": [
    { "rule": "<rule name>", "scope": "<request|response>",
      "action": "<redacted|blocked>" }
  ],
  "holdId": "<UUID or null>",
  "proxyVersion": "<semver>"
}

8.1. Localhost Proxy

The localhost proxy is the simplest deployment: a single binary running on the developer's machine alongside the AI client. It binds to 127.0.0.1:8787 (configurable), supports HTTP and MCP stdio interception, and uses a local SQLite database for the nonce cache, revocation cache, and audit log. Policy is loaded from a YAML file at ~/.aip/policy.yaml.¶

This mode is suitable for individual developers and is the recommended starting point for adopting AIP. A Go proxy implementation is available at the working group's GitHub repository.¶

8.2. Kubernetes Sidecar

For production deployments, the AIP Proxy runs as a sidecar container in the same pod as the agent container. Outbound tool-server traffic is redirected through the proxy on port 15001 via iptables REDIRECT rules. Policy is loaded from a Kubernetes ConfigMap or Secret. Storage uses Redis for the distributed nonce cache and a PersistentVolumeClaim for the audit log. A Prometheus metrics endpoint is exposed at /metrics.¶

Exported metrics include: aip_calls_total, aip_denials_total, aip_holds_total, and aip_verification_latency_seconds.¶

8.3. Enterprise Federation

In enterprise environments the AIP Proxy integrates with existing identity infrastructure:¶

• OIDC mapping: The registry maintains a table mapping Agent IDs to OIDC subject claims. When a tool server requires an OIDC token, the proxy can present one on the agent's behalf after AIP verification passes, without the agent holding OIDC credentials directly.¶

• SPIFFE/mTLS: The proxy is issued a SPIFFE SVID and uses it to establish mTLS connections to tool servers, providing transport- layer mutual authentication in addition to AIP application-layer identity.¶

• Central policy management: AgentPolicy files are managed via a policy API with versioning, rollback, and change auditing, rather than local YAML files.¶

9.1. Cryptographic Algorithm

AIP is built with the idea of using Ed25519 [RFC8032] for all signatures. Ed25519 was chosen because it is fast (sign and verify in under 1ms), produces short keys and signatures (32 and 64 bytes respectively), is deterministic (no per-signature randomness required), and has broad library support across all major languages and platforms. However other cryptographic algorithms should be supported.¶

All Ed25519 operations MUST use a constant-time
implementation to prevent timing side-channels (Section 9.3).

9.2. Prompt Injection Resistance

The principal threat model for AIP at Layer 2 is the prompt injection attack: malicious content in the agent's context causes it to attempt tool calls outside its intended scope.¶

The allowlist in the AgentPolicy prevents the agent from reaching tools it was not provisioned for, regardless of what the model produces. The "block" action provides an unconditional deny for specified tools that cannot be overridden by any model output, injected prompt, or delegation claim. The proxy operates entirely outside the model's trust boundary; the model cannot influence the proxy's decisions.¶

AIP does not protect against:
o  A compromised agent runtime that routes calls around the proxy;
o  Tool servers that accept calls from sources other than the proxy.

9.3. Transport Security

All communication between AIP Proxies and the AIP Registry MUST use TLS 1.3 [RFC8446]. Server certificates MUST be validated against the system trust store. Certificate pinning is RECOMMENDED for registry connections in production deployments.¶

Proxy-to-tool-server connections MUST use TLS 1.2 [RFC5246] or later. TLS 1.3 is RECOMMENDED.¶

9.4. Private Key Storage

Agent private keys MUST be stored in a secure key store. In order of preference:¶

1. Hardware Security Module (HSM) or Trusted Platform Module (TPM); 2. OS keychain (macOS Keychain, Windows DPAPI, Linux Secret Service); 3. Encrypted file with passphrase derived using Argon2id [RFC9106].¶

Private keys MUST NOT be stored in environment variables, plaintext configuration files, or source code repositories.¶

9.5. Registry Trust

A proxy MUST be configured with an explicit list of trusted registry hostnames and the TLS certificate fingerprint (or CA) for each. Agent Records from registries not on this list MUST be rejected.¶

9.6. Nonce Cache Sizing

The nonce cache must be large enough to hold all unique nonces generated within the TTL window (600 seconds). Implementations MUST use a bounded cache with LRU eviction and MUST NOT silently drop old nonces without ensuring they are outside the TTL window.¶

9.7. Revocation Latency

The proxy's revocation cache has a maximum refresh interval of 60 seconds (Section 6.1). In the worst case, a revoked agent can continue making calls for up to 60 seconds after revocation. Deployments with stricter requirements SHOULD subscribe to the registry's SSE revocation stream (Section 5.3) and invalidate the cache on receipt of a revocation event.¶

9.8. Denial-of-Service

The proxy adds a small amount of latency to every tool call (one cache lookup and one key verification). To prevent the proxy from becoming an availability bottleneck:¶

• Agent Record lookups MUST be served from the local cache except on cache miss or invalidation;¶

• The proxy SHOULD enforce a per-agent call rate limit to prevent a runaway agent from flooding the registry with cache-miss lookups.¶

10.1. Agent Record Visibility

Agent Records are visible to any party that can query the registry. Principals SHOULD use non-identifying agent names for agents that handle sensitive workloads.¶

10.2. Audit Log Sensitivity

Audit logs contain Agent IDs, tool names, and argument hashes. While argument values are not logged in plaintext, the combination of Agent ID and tool name may itself be sensitive in some contexts. Audit logs MUST be access-controlled appropriately.¶

10.3. DLP Redaction

When DLP redaction is applied to a response, the proxy modifies the data the agent receives. Operators MUST ensure that redaction does not cause the agent to produce incorrect or harmful downstream actions due to missing context.¶

10.4. Registry Data Retention

Registry operators MUST publish a data retention policy. Agent Records for revoked agents SHOULD be pseudonymized after the minimum retention period required for audit log verification.¶

11.1. HTTP Header Field

This document defines the "AIP-Token" HTTP header field. Registration is requested in the "Permanent Message Header Field Names" registry per [RFC3864]. The header field name is "AIP-Token", applicable protocol is "http", status is "standard", the author/change controller is the IETF, and the specification document is this document (Section 7.1).¶

11.2. Media Type

The media type "application/aip+json" is requested for AIP Tokens serialized as standalone JSON documents. The type name is "application", the subtype name is "aip+json", the required parameter is "version" (value "1"), encoding considerations are UTF-8, and security considerations are described in Section 9.¶