<?xml version="1.0" encoding="utf-8"?>
<!DOCTYPE rfc [ <!ENTITY nbsp "&#160;"> ]>
<rfc xmlns:xi="http://www.w3.org/2001/XInclude" category="std" docName="draft-ganesan-agent-authenticator-00" ipr="trust200902"
     submissionType="IETF" version="3" consensus="true" sortRefs="true" symRefs="true" tocInclude="true" xml:lang="en">
<front>
<title abbrev="Agent Authenticator">Agent Authenticator: Hardware-Bound, Origin-Bound Authentication for AI Agents</title>
<seriesInfo name="Internet-Draft" value="draft-ganesan-agent-authenticator-00"/>
<author fullname="Harish Santhanalakshmi Ganesan" initials="H. S." surname="Ganesan">
<address><email>harishsg99@gmail.com</email></address>
</author>
<date year="2026" month="November" day="15"/>
<area>Security</area>
<keyword>AI agents</keyword><keyword>authentication</keyword><keyword>attestation</keyword><keyword>passkeys</keyword>
<abstract>
<t>This document specifies the Agent Authenticator, a mechanism that lets autonomous software agents, such as AI coding agents, authenticate to services with properties comparable to passkeys: keys held in secure hardware, no shared secrets, binding to the service's origin that a manipulated agent cannot override, attested execution environments, and explicit human approval of each agent and of high-risk actions. The same protocol serves agents on personal computers and in cloud environments. It is a profile over existing mechanisms rather than a new identity system: it uses OAuth 2.0 Token Exchange, the Identity Assertion JSON Web Token (JWT) Authorization Grant, and Model Context Protocol (MCP) authorization, and it binds each agent action to a signed, tamper-evident record.</t>
</abstract>
</front>
<middle>
<section anchor="introduction"><name>Introduction</name>
<t>Software agents built on large language models increasingly act on behalf of people when using services such as code hosts, package registries, model APIs, MCP servers and cloud platforms. This document specifies how an agent authenticates to those services with properties comparable to passkeys, without giving the agent any secret it could leak. It defines one new component, the Agent Authenticator, along with its data formats, ceremonies and bindings to existing OAuth, MCP, git and cloud mechanisms.</t>
<section anchor="goals-and-non-goals"><name>Goals and Non-Goals</name>
<t>The mechanism is designed to meet the following goals:</t>
<ol>
<li>The agent process does not hold private keys or any other secret that can be used to obtain tokens. It requests signatures from a separate authenticator.</li>
<li>A credential can be used only with the relying party for which it was created. The origin is taken from the network connection rather than from the agent's output, so an agent that has been manipulated through prompt injection cannot direct a credential to another site.</li>
<li>Agent keys are generated and held in secure hardware or in an attested cloud key service, and relying parties can learn how well a key is protected through a normalized assurance level.</li>
<li>A named person approves each agent credential, and each high-risk action, using a passkey. The person's device need not be the device on which the agent runs.</li>
<li>Deployment does not require a new identity system. Enterprise identity providers remain the policy authority, and services accept agent credentials through existing OAuth grants.</li>
<li>Each action that the authenticator authorizes is recorded in a signed, tamper-evident log that can be exported to existing security monitoring systems.</li>
<li>The same protocol is used whether the agent runs on a personal computer or in a cloud environment; only the implementation of the authenticator differs.</li>
</ol>
<t>The following are out of scope: replacing OAuth, OpenID Connect or WebAuthn; agent-to-agent messaging; and assessing whether an agent's output is correct.</t>
</section>
<section anchor="motivation"><name>Motivation</name>
<t>Coding agents today authenticate with the same bearer secrets humans use (PATs, API keys, npm tokens, cloud keys), stored where the model, its subprocesses and any malware can read them. The incidents below, all reported in 2025 or 2026, illustrate the resulting failures.</t>
<section anchor="evidence-from-real-incidents"><name>Evidence from Real Incidents</name>
<table><thead><tr><th>Incident (date)</th><th>What happened</th><th>Root cause</th><th>Property</th></tr></thead><tbody><tr><td>GitHub MCP prompt injection (May 2025)</td><td>A malicious issue in a public repo made an agent pull private repo data and publish it in a public PR</td><td>Broad PAT covering every repo the user can reach</td><td>P8 mandate scopes repos and actions; P5 step-up for cross-repo writes</td></tr><tr><td>Comment and Control (Apr 2026)</td><td>One PR title hijacked several CI coding agents, which leaked repo secrets back through GitHub itself</td><td>Secrets ambient in the agent's environment</td><td>P2 model never holds secrets; P8 review agents get no write or secret scopes</td></tr><tr><td>Six coding-agent exploits (2025-2026)</td><td>Branch names stole tokens; repo settings files skipped trust dialogs; sandboxes were escaped</td><td>Controls lived inside the agent and could be bypassed from repo content</td><td>AA enforces policy outside the agent; P9 every agent registered and tied to a human</td></tr><tr><td>Nx s1ngularity (Aug 2025)</td><td>Malicious package versions harvested 2,349 credentials from 1,079 developer machines, using local AI CLIs to hunt for secrets</td><td>Tokens and keys readable on disk by any local process</td><td>P1, P2 nothing to steal; AA signs only for processes it launched</td></tr><tr><td>Shai-Hulud worm waves (Sep 2025 - Oct 2026)</td><td>Self-propagating worm steals npm and GitHub tokens and republishes itself into every package the victim maintains</td><td>Long-lived publish tokens usable from any machine</td><td>P7 no exportable credentials; publish requires <tt>ha</tt></td></tr></tbody></table>
<t>Sources: <xref target="INVARIANT"/>, <xref target="CSA-CI"/>, <xref target="VB-EXPLOITS"/>, <xref target="NX-LEAK"/> and <xref target="SHAI-HULUD"/>.</t>
</section>
<section anchor="evidence-from-credential-data"><name>Evidence from Credential Data</name>
<ul><li>A study by Astrix Security <xref target="ASTRIX"/> of 5,205 open-source MCP servers found 88% need credentials, 53% use static API keys or PATs, only 8.5% use OAuth, and 79% pass keys through environment variables.</li><li>GitGuardian <xref target="GG-SPRAWL"/> found 28.65 million newly leaked secrets on public GitHub in 2025, an 81% increase in leaked secrets for AI services, and 24,008 secrets in MCP configuration files, 2,117 of them valid.</li></ul>
</section>
<section anchor="gaps-in-existing-standards"><name>Gaps in Existing Standards</name>
<table><thead><tr><th>Standard</th><th>What it gives</th><th>What is missing for agents</th></tr></thead><tbody><tr><td>WebAuthn / passkeys</td><td>Phishing-resistant, hardware-bound keys</td><td>Needs a human gesture on every use; no programmatic enrollment</td></tr><tr><td>OAuth 2.1 / MCP authorization</td><td>Scoped, expiring tokens; standard discovery</td><td>Bearer tokens by default; no binding to a device or attested environment</td></tr><tr><td>ID-JAG / Cross App Access</td><td>IdP decides which apps an agent may reach</td><td>Based on the user's session; no proof of where the agent key lives or that a human approved it</td></tr><tr><td>Workload identity (SPIFFE, WIMSE, cloud federation)</td><td>Secretless identity for cloud workloads</td><td>Cloud-centric; no human approver; weak fit for laptops</td></tr><tr><td>Secret vaults and proxies</td><td>Keep secrets out of the model's context</td><td>Still bearer secrets underneath; no standard format, origin binding or attestation</td></tr></tbody></table>
<t>The author is not aware of an existing standard that combines hardware-bound keys, origin binding that survives prompt injection, a human-signed scope, environment attestation, and a single protocol for local and cloud agents.</t>
</section>
</section>
<section anchor="requirements-language"><name>Requirements Language</name>
<t>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 <xref target="RFC2119"/> <xref target="RFC8174"/> when, and only when, they appear in all capitals, as shown here.</t>
</section>
</section>
<section anchor="terminology-and-architecture"><name>Terminology and Architecture</name>
<t>The architecture involves the roles listed below. Of these, only the Agent Authenticator is introduced by this document.</t>
<table><thead><tr><th>Role</th><th>What it is</th><th>Passkey analog</th></tr></thead><tbody><tr><td>Human approver</td><td>The accountable person who owns or delegates to the agent</td><td>User</td></tr><tr><td>Approver authenticator</td><td>The human's passkey device (phone, laptop, security key)</td><td>Authenticator</td></tr><tr><td>Agent</td><td>The model-driven process (coding agent, its subprocesses and tools)</td><td>None (untrusted caller)</td></tr><tr><td>Agent runtime</td><td>The harness that launches the agent and owns its network connections</td><td>Browser / client</td></tr><tr><td>Agent Authenticator (AA)</td><td>Daemon that holds agent keys in secure hardware, enforces mandates, signs assertions and action records</td><td>Authenticator</td></tr><tr><td>Relying party (RP)</td><td>Service the agent calls: code host, package registry, model API, MCP server, cloud</td><td>Relying party</td></tr><tr><td>Identity provider (IdP)</td><td>Enterprise policy authority, such as a single sign-on provider</td><td>None</td></tr><tr><td>Evidence pipeline</td><td>Local sealed log, SIEM and a SCITT transparency service</td><td>None</td></tr></tbody></table>
<figure anchor="fig-architecture-overview"><name>Architecture overview</name><artwork type="ascii-art"><![CDATA[
                                     +--------------------+
                                     |   Enterprise IdP   |
                                     | policy, revocation |
                                     +---------+----------+
                                               | token exchange
                                               | (ID-JAG)
+-----------+  call  +---------------+ assert. +--------------------+
|   Agent   |------->| Agent runtime |-------->|  Relying parties   |
| (no keys) |        | sets origin   |         |  GitHub, npm, MCP  |
+-----------+        +---+-------^---+         +--------------------+
                    sign |       | signed
                 request v       | assertion
+-----------+  approval  +---------------+   revocation (SSF/CAEP)
|   Human   |----------->|     Agent     |<------------- from IdP
| (passkey) |            | Authenticator |
+-----------+            | keys in HW    |
                         +---------------+
]]></artwork></figure>
<t>The agent can request a signature but never receives a key. The runtime supplies the origin of each request from its own connection state, and only the Agent Authenticator holds keys. Where an IdP is present, it remains the policy authority and can revoke an agent's access.</t>
<t>This document uses the following terms:</t>
<dl newline="false"><dt>Agent credential:</dt><dd>An RP-scoped key pair created and held by the AA for one agent instance.</dd><dt>Mandate:</dt><dd>A human-signed statement of what an agent credential may do, where, and until when.</dd><dt>Environment attestation:</dt><dd>Hardware evidence about where the AA's keys live and what runs there.</dd><dt>Assurance level (AL):</dt><dd>A normalized rating of environment attestation (<xref target="keys-and-assurance-levels"/>).</dd><dt>Action record:</dt><dd>A signed, hash-chained record of one agent action, bound to the credential that authorized it.</dd></dl>
<section anchor="abbreviations"><name>Abbreviations</name>
<dl spacing="compact">
<dt>AA:</dt><dd>Agent Authenticator</dd>
<dt>AL:</dt><dd>assurance level</dd>
<dt>CAEP:</dt><dd>Continuous Access Evaluation Profile</dd>
<dt>COSE:</dt><dd>CBOR Object Signing and Encryption</dd>
<dt>DPoP:</dt><dd>Demonstrating Proof of Possession</dd>
<dt>EAT:</dt><dd>Entity Attestation Token</dd>
<dt>ID-JAG:</dt><dd>Identity Assertion JWT Authorization Grant</dd>
<dt>IdP:</dt><dd>identity provider</dd>
<dt>KMS:</dt><dd>key management service</dd>
<dt>MCP:</dt><dd>Model Context Protocol</dd>
<dt>OCSF:</dt><dd>Open Cybersecurity Schema Framework</dd>
<dt>PAT:</dt><dd>personal access token</dd>
<dt>RAR:</dt><dd>Rich Authorization Requests</dd>
<dt>RATS:</dt><dd>Remote ATtestation procedureS</dd>
<dt>RP:</dt><dd>relying party</dd>
<dt>SCITT:</dt><dd>Supply Chain Integrity, Transparency, and Trust</dd>
<dt>SIEM:</dt><dd>security information and event management</dd>
<dt>SSF:</dt><dd>Shared Signals Framework</dd>
<dt>TEE:</dt><dd>trusted execution environment</dd>
<dt>TPM:</dt><dd>Trusted Platform Module</dd>
</dl>
</section>
</section>
<section anchor="security-properties"><name>Security Properties</name>
<t>The following table compares the security properties of passkeys with those of this mechanism. Properties P8 to P10 have no counterpart in passkeys.</t>
<table><thead><tr><th>#</th><th>Property</th><th>Passkeys</th><th>This specification</th></tr></thead><tbody><tr><td>P1</td><td>No shared secrets</td><td>Public-key challenge-response</td><td>Same; RPs store only public keys</td></tr><tr><td>P2</td><td>Key non-exportable</td><td>Platform or roaming authenticator</td><td>TPM, Secure Enclave, StrongBox, TEE or attested cloud KMS; model process never sees keys</td></tr><tr><td>P3</td><td>Origin binding</td><td>Browser sets origin; key scoped to RP ID</td><td>Runtime sets origin from the TLS connection; key scoped to RP ID; model output can never choose it</td></tr><tr><td>P4</td><td>Attestation</td><td>Optional authenticator attestation</td><td>Required environment attestation, normalized to assurance levels</td></tr><tr><td>P5</td><td>Presence / verification</td><td>UP and UV flags</td><td>MB flag (acting within mandate) and HA flag (human approved this action)</td></tr><tr><td>P6</td><td>Clone detection</td><td>Signature counter</td><td>Signature counter plus action-record chain head in every assertion</td></tr><tr><td>P7</td><td>Sync</td><td>Passkeys may sync</td><td>Agent credentials are never synchronized; in the cloud profile they are destroyed when the task ends</td></tr><tr><td>P8 (new)</td><td>Scoped authority</td><td>None</td><td>Mandate: RP list, actions, resources, expiry, minimum assurance</td></tr><tr><td>P9 (new)</td><td>Accountability</td><td>None</td><td>Every assertion names its human approver and mandate hash</td></tr><tr><td>P10 (new)</td><td>Identity-bound evidence</td><td>None</td><td>Every authorized action yields a signed, hash-chained action record</td></tr></tbody></table>
<t>P3 is the property that most directly addresses prompt injection. If an agent is instructed to send a credential to a server controlled by an attacker, it cannot do so: no credential exists for that origin, and the agent has no way to set the origin itself.</t>
</section>
<section anchor="keys-and-assurance-levels"><name>Keys and Assurance Levels</name>
<t>The mechanism uses four keys. None of them is accessible to the agent process.</t>
<table><thead><tr><th>Key</th><th>Held by</th><th>Created</th><th>Lifetime</th></tr></thead><tbody><tr><td>Approver key</td><td>Human's passkey authenticator</td><td>Normal WebAuthn registration with the IdP or RP</td><td>Long-lived; device-bound passkey RECOMMENDED for approvals</td></tr><tr><td>AA identity key</td><td>AA secure hardware</td><td>AA install or cloud environment boot</td><td>Life of the device or environment</td></tr><tr><td>Agent credential key</td><td>AA secure hardware</td><td>Per agent instance, per RP</td><td>Local: until revoked or mandate expiry. Cloud: one task</td></tr><tr><td>Evidence signing key</td><td>AA secure hardware, separate from credential keys</td><td>With the AA identity key</td><td>Rotated per epoch, forward-secure</td></tr></tbody></table>
<t>Verifiers MUST base policy decisions on the normalized assurance level defined below, rather than on vendor-specific attestation formats.</t>
<table><thead><tr><th>Level</th><th>Meaning</th><th>Example evidence</th></tr></thead><tbody><tr><td>AL3</td><td>Keys in a TEE that also isolates the agent's code; runtime measured</td><td>AWS Nitro, Intel TDX, AMD SEV-SNP, Arm CCA attestation report</td></tr><tr><td>AL2</td><td>Keys in dedicated secure hardware; host software not isolated</td><td>TPM 2.0 key attestation, Apple App Attest, Android StrongBox key attestation</td></tr><tr><td>AL1</td><td>Keys in a platform keystore without hardware isolation, or cloud workload identity only</td><td>Android TEE-only keystore, cloud instance identity document, SPIFFE SVID</td></tr><tr><td>AL0</td><td>Software key</td><td>None</td></tr></tbody></table>
<t>The AA converts vendor evidence into an Entity Attestation Token <xref target="RFC9711"/> carrying <tt>al</tt>, platform, key-protection claims and a freshness nonce. A RATS <xref target="RFC9334"/> verifier appraises it; a separate EAT profile is expected to standardize the mapping.</t>
<t>A passkey that is synchronized between devices authenticates the person but does not identify the device that was used. Approvals that require device binding (<xref target="step-up"/>) MUST use a device-bound authenticator or carry a separate platform device attestation.</t>
</section>
<section anchor="data-formats"><name>Data Formats</name>
<t>All structures are JSON Web Signatures <xref target="RFC7515"/> for HTTP and OAuth bindings, with a COSE <xref target="RFC9052"/> encoding for constrained and SSH bindings.</t>
<section anchor="mandate"><name>Mandate</name>
<t>The mandate is signed by the human approver.</t>
<sourcecode type="json"><![CDATA[
{
  "typ": "agent-mandate+jwt",
  "iss": "<IdP or approver identifier>",
  "sub": "<human id>",
  "agent_cred": { "jkt": "<agent key thumbprint, RFC 7638>" },
  "aa": { "id": "<AA identity key thumbprint>", "min_al": "AL2" },
  "rps": ["code.example", "registry.example", "mcp.example"],
  "authorization_details": [
    { "type": "git", "actions": ["push"],
      "locations": ["code.example/acme/api"],
      "refs": ["agent/fix-*"] },
    { "type": "mcp", "tools": ["jira.read", "jira.comment"] }
  ],
  "step_up": ["merge", "publish", "secrets.read", "delete"],
  "delegable": false,
  "exp": 1762480000
}
]]></sourcecode>
<t>The <tt>jkt</tt> member is a JWK Thumbprint <xref target="RFC7638"/>. The approver signs the mandate through a WebAuthn <xref target="WebAuthn"/> assertion whose challenge is <tt>SHA-256(mandate)</tt>. The <tt>authorization_details</tt> member reuses <xref target="RFC9396"/> so IdPs and RPs evaluate it with existing policy engines.</t>
</section>
<section anchor="agentdata"><name>AgentData</name>
<t>AgentData is the signed core of every assertion, analogous to WebAuthn authenticatorData.</t>
<sourcecode type="json"><![CDATA[
{
  "rp_id": "code.example",
  "origin": "https://api.code.example",
  "challenge": "<RP nonce>",
  "cred_id": "<agent credential id>",
  "flags": { "mb": true, "ha": false, "ev": true },
  "mandate_hash": "<SHA-256 of mandate>",
  "al": "AL2",
  "counter": 412,
  "evidence_head": "<hash of latest action record>",
  "action": { "type": "git.push",
              "target": "acme/api@agent/fix-123" }
}
]]></sourcecode>
<t>The <tt>origin</tt> and <tt>rp_id</tt> members are filled by the runtime from the live connection. The AA MUST reject any request where they do not match a credential's scope.</t>
</section>
<section anchor="assertion"><name>Assertion</name>
<t>The assertion is AgentData signed by the agent credential key. When <tt>ha</tt> is set, it also carries the approver's WebAuthn assertion over <tt>SHA-256(AgentData)</tt> for this specific action.</t>
</section>
<section anchor="attestation-statement"><name>Attestation Statement</name>
<t>Used at registration only: an EAT <xref target="RFC9711"/> containing the AA identity key, the agent credential public key, <tt>al</tt>, platform evidence (raw vendor statement, optional), the RP-provided nonce, and the environment measurement where AL3 applies.</t>
</section>
<section anchor="agentdata-claim-definitions"><name>AgentData Claim Definitions</name>
<table><thead><tr><th>Claim</th><th>Type</th><th>Requirement</th><th>Meaning</th></tr></thead><tbody><tr><td><tt>rp_id</tt></td><td>string</td><td>REQUIRED</td><td>RP ID the credential is scoped to; set by the runtime</td></tr><tr><td><tt>origin</tt></td><td>string</td><td>REQUIRED</td><td>Origin of the live TLS connection; never set by the agent</td></tr><tr><td><tt>challenge</tt></td><td>string</td><td>REQUIRED</td><td>Fresh nonce from the RP, at least 128 bits</td></tr><tr><td><tt>cred_id</tt></td><td>string</td><td>REQUIRED</td><td>Identifier of the agent credential</td></tr><tr><td><tt>flags</tt></td><td>object</td><td>REQUIRED</td><td><tt>mb</tt> within mandate, <tt>ha</tt> human approved, <tt>ev</tt> action record written</td></tr><tr><td><tt>mandate_hash</tt></td><td>string</td><td>REQUIRED</td><td>SHA-256 of the mandate in force</td></tr><tr><td><tt>al</tt></td><td>string</td><td>REQUIRED</td><td>Assurance level, AL0 to AL3</td></tr><tr><td><tt>counter</tt></td><td>integer</td><td>REQUIRED</td><td>Strictly increasing per credential</td></tr><tr><td><tt>action</tt></td><td>object</td><td>REQUIRED</td><td>Action type and target the signature authorizes</td></tr><tr><td><tt>evidence_head</tt></td><td>string</td><td>OPTIONAL</td><td>Hash of the latest action record (<xref target="action-records"/>)</td></tr></tbody></table>
</section>
</section>
<section anchor="ceremonies"><name>Ceremonies</name>
<t>This section defines five ceremonies. Of these, only registration and step-up require a person to take part.</t>
<section anchor="registration"><name>Registration</name>
<ol><li>The runtime asks the RP (or the IdP acting for it) to register an agent credential. The RP returns a nonce and its RP ID.</li><li>The AA creates a key pair scoped to that RP ID in secure hardware.</li><li>The AA builds the attestation statement (<xref target="attestation-statement"/>) bound to the nonce.</li><li>The AA builds a draft mandate and sends it to the human approver: locally as an OS prompt, remotely as a push or QR code to their phone.</li><li>The human reviews a plain-language rendering of the mandate and signs it with their passkey.</li><li>The RP verifies the attestation, the approver signature and the mandate, then stores the agent credential public key linked to the human's account, separately revocable.</li></ol>
<figure anchor="fig-registration"><name>Registration</name><artwork type="ascii-art"><![CDATA[
Human            Agent             Agent           Relying
approver         Authenticator     runtime         party / IdP
  |                  |                |                |
  |                  |                | 1. register    |
  |                  |                |--------------->|
  |                  |                | 2. nonce, RP ID|
  |                  |                |<---------------|
  |                  | 3. create key  |                |
  |                  |<---------------|                |
  |                  |--+ 4. HW key,  |                |
  |                  |<-+ attestation |                |
  | 5. draft mandate |                |                |
  |<-----------------|                |                |
  | 6. passkey signs |                |                |
  |----------------->|                |                |
  |                  | 7. key, attest.|                |
  |                  |    mandate     |                |
  |                  |--------------->|                |
  |                  |                | 8. submit      |
  |                  |                |--------------->|
  |                  |                |                |--+ 9. verify
  |                  |                |                |<-+ and store
]]></artwork></figure>
</section>
<section anchor="authentication"><name>Authentication</name>
<ol><li>The agent calls an RP through the runtime. The RP returns a challenge.</li><li>The runtime fills <tt>rp_id</tt> and <tt>origin</tt> from the connection and forwards the request to the AA.</li><li>The AA checks the mandate (RP allowed, action allowed, not expired, AL still met), writes an action record, and signs AgentData.</li><li>The RP verifies the assertion (<xref target="relying-party-verification-procedure"/>), then issues a short-lived access token bound to the agent key with DPoP <xref target="RFC9449"/>.</li></ol>
</section>
<section anchor="step-up"><name>Step-Up</name>
<ol><li>The RP requires <tt>ha</tt> for an action in the mandate's <tt>step_up</tt> list, or the AA requires it by local policy.</li><li>The AA sends the action summary to the approver's authenticator (same device, or phone via push or hybrid transport).</li><li>The human approves with a passkey assertion over <tt>SHA-256(AgentData)</tt>.</li><li>The AA attaches it and sets <tt>ha</tt>. The approval covers exactly that action and cannot be replayed.</li></ol>
<t>This ceremony is the same for local and cloud agents, because the approval is always made on the person's own authenticator.</t>
<figure anchor="fig-authentication-with-step-up"><name>Authentication with step-up</name><artwork type="ascii-art"><![CDATA[
 Agent    Runtime       Authenti-     Human        Relying
   |         |          cator         (phone)      party
   |         |             |             |            |
   | 1. call |             |             |            |
   |-------->| 2. request  |             |            |
   |         |---------------------------------------->|
   |         |                     3. challenge        |
   |         |<----------------------------------------|
   |         | 4. sign +   |             |            |
   |         |    origin   |             |            |
   |         |------------>|             |            |
   |         |             |--+ 5. mandate            |
   |         |             |<-+   check  |            |
   |         |             | 6. approve? |            |  step-up:
   |         |             |------------>|            |  risky
   |         |             | 7. passkey  |            |  actions
   |         |             |<------------|            |  only
   |         | 8. signed   |             |            |
   |         |<------------|             |            |
   |         | 9. assertion                            |
   |         |---------------------------------------->|
   |         |               10. DPoP-bound token      |
   |         |<----------------------------------------|
]]></artwork></figure>
<t>Steps 6 and 7 apply only to actions that require approval, such as merging, publishing or deleting. For other actions, the AA signs on the basis of the mandate alone.</t>
</section>
<section anchor="sub-delegation"><name>Sub-Delegation</name>
<t>An agent MAY hand a narrower task to another agent only if <tt>delegable</tt> is true. The parent AA signs a child mandate that MUST be a subset of the parent (RPs, actions, expiry) and MUST NOT lower <tt>min_al</tt>. Every hop's credential MUST be attested; the chain travels with the assertion and is verifiable offline.</t>
</section>
<section anchor="revocation"><name>Revocation</name>
<ul><li>Any of these MUST revoke an agent credential: mandate expiry, human or administrator revocation, IdP session revocation, failed re-attestation, broken action-record chain.</li><li>The AA is an OpenID Shared Signals Framework <xref target="SSF"/> receiver: a CAEP session-revoked or credential-change event from the IdP destroys the matching keys.</li><li>The AA is also a Shared Signals transmitter: it emits events when attestation fails or tampering is detected.</li><li>RPs MAY check credential status via a token status list for offline verification.</li></ul>
</section>
<section anchor="relying-party-verification-procedure"><name>Relying Party Verification Procedure</name>
<t>On receiving an assertion, the RP MUST perform these steps in order and MUST reject the request if any step fails:</t>
<ol><li>Look up the agent credential by <tt>cred_id</tt>. Reject if unknown or revoked.</li><li>Verify that <tt>rp_id</tt> equals the RP's own RP ID and <tt>origin</tt> is an origin the RP serves.</li><li>Verify that <tt>challenge</tt> equals a nonce the RP issued, is unexpired, and has not been used before.</li><li>Verify the signature over AgentData with the stored credential public key.</li><li>Verify that <tt>counter</tt> is greater than the stored value, then store it. A non-increasing counter indicates a cloned credential; the RP MUST revoke it.</li><li>Verify that <tt>mandate_hash</tt> matches the mandate stored at registration and that the mandate has not expired.</li><li>Verify that <tt>action</tt> is permitted by the mandate's <tt>authorization_details</tt>.</li><li>Verify that <tt>al</tt> is at least the mandate's <tt>min_al</tt> and the RP's own policy minimum.</li><li>If the action is in the mandate's <tt>step_up</tt> list or RP policy requires approval, verify that <tt>ha</tt> is set and that the attached WebAuthn assertion is a valid signature by the registered approver key over <tt>SHA-256(AgentData)</tt>.</li><li>Only then issue a short-lived access token bound to the agent key with DPoP <xref target="RFC9449"/>, or perform the action.</li></ol>
<t>An RP that delegates policy to an IdP (<xref target="protocol-bindings"/>) MAY perform steps 6 to 9 by passing the assertion to the IdP in token exchange, but MUST still perform steps 1 to 5.</t>
</section>
</section>
<section anchor="protocol-bindings"><name>Protocol Bindings</name>
<t>This section describes how the mechanism maps onto protocols that relying parties already support. In most cases the agent itself needs no changes, because the runtime and the AA handle authentication on its behalf.</t>
<table><thead><tr><th>Binding</th><th>Where it plugs in</th><th>How the assertion is used</th></tr></thead><tbody><tr><td>MCP</td><td>MCP authorization <xref target="MCP-AUTH"/> and its DPoP and workload identity extensions</td><td>AA acts as the MCP client's auth provider; tokens are DPoP-bound to the agent key</td></tr><tr><td>Enterprise IdP</td><td>ID-JAG <xref target="ID-JAG"/>: token exchange <xref target="RFC8693"/> then JWT bearer grant <xref target="RFC7523"/></td><td>The user's identity assertion is the subject token and the agent assertion is the actor token; the IdP applies its own policy and issues the ID-JAG</td></tr><tr><td>Agent gateways</td><td>Gateway authorization on each tool call</td><td>Gateway verifies the assertion, reads <tt>al</tt>, <tt>ha</tt> and mandate hash, asks the IdP for the decision</td></tr><tr><td>Generic OAuth APIs</td><td>JWT bearer or token exchange with authorization details <xref target="RFC9396"/></td><td>RP swaps the assertion for a short-lived, key-bound access token</td></tr><tr><td>Git over HTTPS</td><td>git credential helper</td><td>Helper asks the AA for a token per remote; no PATs on disk</td></tr><tr><td>Git over SSH</td><td>ssh-agent protocol with a key type modeled on OpenSSH FIDO keys</td><td>AA acts as an ssh-agent; signatures carry flags and mandate hash</td></tr><tr><td>Commit signing</td><td>SSH or Sigstore-style signing</td><td>Commits signed by the agent key; trailer names approver and mandate hash</td></tr><tr><td>Cloud</td><td>Workload identity federation</td><td>AA or IdP issues a token the cloud trusts; no static cloud keys</td></tr><tr><td>Package registries</td><td>Trusted publishing</td><td>Publish requires <tt>ha</tt>; provenance records agent identity</td></tr><tr><td>Agent hooks</td><td>Pre-tool-use hooks in coding agents</td><td>Hook routes risky actions to step-up before they run</td></tr></tbody></table>
<t>Some relying parties support none of these mechanisms. For such an RP, the AA MAY fetch a legacy token through a broker, but MUST record <tt>al</tt> as no higher than AL1 and MUST NOT expose the token to the model process.</t>
</section>
<section anchor="action-records"><name>Action Records</name>
<t>Action records are written and signed by the AA, using keys held in the same secure hardware as the agent's credentials. The agent cannot create or modify them.</t>
<section anchor="what-is-recorded"><name>What Is Recorded</name>
<t>The AA records the following types of event:</t>
<dl newline="false"><dt>Action events:</dt><dd>Every tool call or API call the AA signs or refuses, one record per action.</dd><dt>Identity events:</dt><dd>Registration, mandate signed, step-up approved or denied, revocation, re-attestation result.</dd><dt>Local actions:</dt><dd>File edits and shell commands reported by the runtime, labeled as runtime-reported.</dd><dt>Integrity events:</dt><dd>AA start and stop, heartbeats, chain verification failures.</dd><dt>OS events (optional profile):</dt><dd>Raw process, file and network activity, for deployments that need forensic depth.</dd></dl>
</section>
<section anchor="record-generation"><name>Record Generation</name>
<t>The AA writes one record per agent action at the moment it signs or refuses it. Because every request that needs a credential passes through the AA, the AA observes every action that the agent takes against a relying party.</t>
<sourcecode type="json"><![CDATA[
{
  "seq": 1042,
  "time": "2026-11-06T14:03:22Z",
  "agent": { "cred": "<agent key thumbprint>",
             "name": "coding-agent@laptop-17" },
  "human": "priya@acme.example",
  "mandate_hash": "<SHA-256 of mandate>",
  "action": {
    "kind": "mcp.tool_call",
    "rp": "mcp.example",
    "tool": "jira.comment",
    "target": "PROJ-142",
    "args_hash": "sha256:<hash of arguments>"
  },
  "decision": "allowed",
  "flags": { "mb": true, "ha": false },
  "al": "AL2",
  "source": "daemon",
  "parent_seq": 1039,
  "prev_hash": "<hash of record 1041>",
  "sig": "<signature by evidence key>"
}
]]></sourcecode>
<table><thead><tr><th>Action</th><th>How the AA sees it</th><th>source</th></tr></thead><tbody><tr><td>MCP tool call to a remote server</td><td>AA is the MCP client's auth provider and signs each call</td><td>daemon</td></tr><tr><td>API call (code host, registry, cloud, model API)</td><td>AA signs the request or the token exchange</td><td>daemon</td></tr><tr><td>git push, commit signing</td><td>AA acts as git credential helper and ssh-agent</td><td>daemon</td></tr><tr><td>Step-up request, approval or denial</td><td>AA runs the ceremony</td><td>daemon</td></tr><tr><td>Action refused as outside the mandate</td><td>AA refuses to sign and records why</td><td>daemon</td></tr><tr><td>Local tool (file edit, shell command, local MCP server)</td><td>Runtime reports it to the AA through its hook</td><td>runtime</td></tr></tbody></table>
<t>The <tt>decision</tt> member takes one of the following values: <tt>allowed</tt>, <tt>step_up_approved</tt>, <tt>step_up_denied</tt>, <tt>denied_out_of_scope</tt>, <tt>denied_origin_mismatch</tt>, <tt>denied_revoked</tt>.</t>
<t>The <tt>parent_seq</tt> member ties each action to the one that led to it, so one task, including sub-agent hand-offs, can be reconstructed as a single sequence of actions, from the request that started the task to its last tool call.</t>
<ul><li>Records are signed only by the AA.</li><li>Arguments are stored as a hash by default; full arguments stay in the local log.</li><li>Runtime-reported records MUST carry <tt>source: runtime</tt>.</li><li>The AA MUST emit a signed heartbeat at a fixed interval, so a stopped AA leaves a provable gap.</li></ul>
</section>
<section anchor="tamper-evidence"><name>Tamper Evidence</name>
<ol><li>Each record carries the hash of the previous record, so edits, deletions and reordering are detectable.</li><li>The evidence key is held in secure hardware and rotates per epoch (forward-secure), so a later compromise cannot forge past records.</li><li>The <tt>evidence_head</tt> member of AgentData commits the chain to each RP, so RPs hold independent witnesses.</li><li>At each epoch, and on every <tt>ha</tt> event, the AA registers a signed statement with the chain head on a SCITT transparency service <xref target="RFC9943"/>.</li></ol>
</section>
<section anchor="siem-delivery"><name>SIEM Delivery</name>
<t>Records are emitted as OCSF <xref target="OCSF"/> events so common SIEMs ingest them without a custom parser. Tool and API calls map to API Activity (6003), authentication and step-up to Authentication (3002), and mandate grants and revocations to Account Change (3001).</t>
</section>
</section>
<section anchor="deployment-profiles"><name>Deployment Profiles</name>
<t>The protocol does not change between the two profiles below. What differs is how the AA is implemented and which assurance level it can reach.</t>
<table><thead><tr><th>Aspect</th><th>Local profile (developer machine)</th><th>Cloud profile (sandbox, VM, enclave)</th></tr></thead><tbody><tr><td>AA form</td><td>OS daemon in a separate privilege domain from the agent</td><td>Sidecar, enclave key service, or attested cloud KMS</td></tr><tr><td>Key storage</td><td>TPM 2.0, Secure Enclave</td><td>Confidential-computing enclave, vTPM, or KMS with attestation-gated release</td></tr><tr><td>Typical AL</td><td>AL2 (AL1 without secure hardware)</td><td>AL3 in confidential computing, AL1 with instance identity only</td></tr><tr><td>Human approval</td><td>Same device or phone</td><td>Always remote: phone via push or hybrid flow</td></tr><tr><td>Credential lifetime</td><td>Until mandate expiry or revocation</td><td>One task; keys destroyed when the environment ends</td></tr><tr><td>Key isolation rule</td><td>AA MUST run outside the agent's user or sandbox and sign only for processes it launched</td><td>AA MUST release keys only after attestation matches the expected measurement</td></tr><tr><td>Hand-off</td><td>A session moving to the cloud gets a new credential under a child mandate; it never copies the local key</td><td>Same rule in reverse</td></tr></tbody></table>
</section>
<section anchor="implementation-status" removeInRFC="true"><name>Implementation Status</name>
<t>This section records the status of known implementations of the protocol defined by this specification at the time of posting of this Internet-Draft, and is based on a proposal described in <xref target="RFC7942"/>. The description of implementations in this section is intended to assist the IETF in its decision processes in progressing drafts to RFCs. Please note that the listing of any individual implementation here does not imply endorsement by the IETF. Furthermore, no effort has been spent to verify the information presented here that was supplied by IETF contributors. This is not intended as, and must not be construed to be, a catalog of available implementations or their features. Readers are advised to note that other implementations may exist.</t>
<t>According to <xref target="RFC7942"/>, "this will allow reviewers and working groups to assign due consideration to documents that have the benefit of running code, which may serve as evidence of valuable experimentation and feedback that have made the implemented protocols more mature. It is up to the individual working groups to use this information as they see fit".</t>
<t>At the time of writing, there are no known implementations. A reference implementation of the Agent Authenticator for Linux with TPM 2.0, for macOS with the Secure Enclave, and for one confidential-computing environment is planned.</t>
</section>
<section anchor="security-considerations"><name>Security Considerations</name>
<t>The mechanism limits what a compromised agent can do and makes tampering with records detectable. It does not protect against a fully compromised host.</t>
<table><thead><tr><th>Threat</th><th>Mitigation</th><th>Residual risk</th></tr></thead><tbody><tr><td>Prompt injection sends agent to attacker origin</td><td>P3 origin binding; no credential exists for that origin</td><td>Injection can still misuse allowed RPs within the mandate</td></tr><tr><td>Injection abuses an allowed RP</td><td>Narrow mandates, step-up for destructive actions, action records</td><td>Damage up to mandate scope</td></tr><tr><td>Token theft from agent memory or logs</td><td>Model never holds keys; tokens DPoP-bound and short-lived</td><td>A token used from the same host before expiry</td></tr><tr><td>Same-user malware borrows the AA</td><td>AA in separate privilege domain, process-lineage check, <tt>ha</tt> for risky actions</td><td>Local administrator can subvert the AA</td></tr><tr><td>Administrator deletes or edits records</td><td>Hash chain, forward-secure hardware signing, RP-held heads, SCITT receipts</td><td>Later events may be missing, but the gap is provable</td></tr><tr><td>Approval fatigue</td><td>Plain-language rendering of exactly what is signed; rate limits; risk-tiered <tt>ha</tt></td><td>Humans can still approve bad requests</td></tr><tr><td>Replay of approvals</td><td>Approval signs <tt>SHA-256(AgentData)</tt> including RP challenge and counter</td><td>None known</td></tr><tr><td>Stale attestation</td><td>Fresh nonce at registration; periodic and per-<tt>ha</tt> re-attestation</td><td>Window between re-attestations</td></tr><tr><td>Over-broad sub-delegation</td><td>Child mandate must be a subset; <tt>min_al</tt> cannot drop</td><td>Depends on correct policy authoring</td></tr></tbody></table>
<t>Three limitations should be noted. The record mechanisms detect tampering but do not prevent it. Guarantees on personal computers are weaker than those available in attested enclaves. Finally, assurance ultimately depends on the attestation roots of hardware and cloud vendors.</t>
</section>
<section anchor="privacy-considerations"><name>Privacy Considerations</name>
<ul><li>Hardware attestation is linkable across RPs. RPs SHOULD receive only the normalized <tt>al</tt>, not raw vendor evidence, and AAs use per-RP credential keys (P3).</li><li>Action records contain sensitive data (file paths, commands, tool arguments). The local log stores full records; the SIEM receives configurable redaction; the transparency service receives only hashes.</li><li>Unlinkable presentation of <tt>al</tt>, for example using anonymous credentials such as BBS signatures or TPM Direct Anonymous Attestation, is left for future work.</li></ul>
</section>
<section anchor="iana-considerations"><name>IANA Considerations</name>
<section anchor="iana-media-types"><name>Media Types</name>
<t>IANA is requested to register the following media types <xref target="RFC6838"/> in the "Media Types" registry <xref target="IANA.MEDIA"/>.</t>
<section anchor="iana-mandate-type"><name>application/agent-mandate+jwt</name>
<dl spacing="compact">
<dt>Type name:</dt><dd>application</dd>
<dt>Subtype name:</dt><dd>agent-mandate+jwt</dd>
<dt>Required parameters:</dt><dd>N/A</dd>
<dt>Optional parameters:</dt><dd>N/A</dd>
<dt>Encoding considerations:</dt><dd>binary; a JWT is a sequence of base64url-encoded values separated by period characters</dd>
<dt>Security considerations:</dt><dd>See <xref target="security-considerations"/> of this document</dd>
<dt>Interoperability considerations:</dt><dd>N/A</dd>
<dt>Published specification:</dt><dd><xref target="mandate"/> of this document</dd>
<dt>Applications that use this media type:</dt><dd>Agent Authenticators, relying parties and identity providers that process agent mandates</dd>
<dt>Fragment identifier considerations:</dt><dd>N/A</dd>
<dt>Additional information:</dt><dd>Magic number(s): N/A; File extension(s): N/A; Macintosh file type code(s): N/A</dd>
<dt>Person and email address to contact for further information:</dt><dd>Harish Santhanalakshmi Ganesan, harishsg99@gmail.com</dd>
<dt>Intended usage:</dt><dd>COMMON</dd>
<dt>Restrictions on usage:</dt><dd>none</dd>
<dt>Author:</dt><dd>Harish Santhanalakshmi Ganesan</dd>
<dt>Change controller:</dt><dd>IETF</dd>
</dl>
</section>
<section anchor="iana-assertion-type"><name>application/agent-assertion+jwt</name>
<dl spacing="compact">
<dt>Type name:</dt><dd>application</dd>
<dt>Subtype name:</dt><dd>agent-assertion+jwt</dd>
<dt>Required parameters:</dt><dd>N/A</dd>
<dt>Optional parameters:</dt><dd>N/A</dd>
<dt>Encoding considerations:</dt><dd>binary; a JWT is a sequence of base64url-encoded values separated by period characters</dd>
<dt>Security considerations:</dt><dd>See <xref target="security-considerations"/> of this document</dd>
<dt>Interoperability considerations:</dt><dd>N/A</dd>
<dt>Published specification:</dt><dd><xref target="assertion"/> of this document</dd>
<dt>Applications that use this media type:</dt><dd>Agent Authenticators, relying parties and identity providers that process agent assertions</dd>
<dt>Fragment identifier considerations:</dt><dd>N/A</dd>
<dt>Additional information:</dt><dd>Magic number(s): N/A; File extension(s): N/A; Macintosh file type code(s): N/A</dd>
<dt>Person and email address to contact for further information:</dt><dd>Harish Santhanalakshmi Ganesan, harishsg99@gmail.com</dd>
<dt>Intended usage:</dt><dd>COMMON</dd>
<dt>Restrictions on usage:</dt><dd>none</dd>
<dt>Author:</dt><dd>Harish Santhanalakshmi Ganesan</dd>
<dt>Change controller:</dt><dd>IETF</dd>
</dl>
</section>
</section>
<section anchor="iana-oauth-uri"><name>OAuth URI</name>
<t>IANA is requested to register the following value in the "OAuth URI" registry <xref target="IANA.OAUTH"/>. The agent assertion is used as the <tt>actor_token</tt> in token exchange <xref target="RFC8693"/>, so that the agent is identified as acting on behalf of the user named by the subject token.</t>
<dl spacing="compact">
<dt>URN:</dt><dd>urn:ietf:params:oauth:token-type:agent-assertion</dd>
<dt>Common Name:</dt><dd>Agent assertion</dd>
<dt>Change Controller:</dt><dd>IETF</dd>
<dt>Specification Document:</dt><dd><xref target="assertion"/> of this document</dd>
</dl>
</section>
<section anchor="iana-jwt-claims"><name>JSON Web Token Claims</name>
<t>IANA is requested to register the following claims in the "JSON Web Token Claims" registry <xref target="IANA.JWT"/>. For each, the Change Controller is IETF and the Specification Document is <xref target="agentdata-claim-definitions"/> of this document.</t>
<table><thead><tr><th>Claim Name</th><th>Claim Description</th></tr></thead><tbody>
<tr><td>agent_cred</td><td>Agent credential, identified by key thumbprint</td></tr>
<tr><td>mandate_hash</td><td>Hash of the mandate in force</td></tr>
<tr><td>al</td><td>Agent environment assurance level</td></tr>
<tr><td>evidence_head</td><td>Hash of the latest action record</td></tr>
</tbody></table>
</section>
<section anchor="iana-new-registries"><name>Agent Authenticator Registry Group</name>
<t>IANA is requested to create a new registry group named "Agent Authenticator" containing the two registries below. Both use the "Specification Required" registration policy <xref target="RFC8126"/>. Designated experts should check that a proposed value is clearly defined, does not duplicate an existing value, and, for assurance levels, states the evidence a verifier needs to assign it.</t>
<section anchor="iana-al-registry"><name>Agent Assurance Levels Registry</name>
<t>Template: Level name; Description; Reference. Initial contents:</t>
<table><thead><tr><th>Level</th><th>Description</th><th>Reference</th></tr></thead><tbody>
<tr><td>AL0</td><td>Software key</td><td>This document</td></tr>
<tr><td>AL1</td><td>Platform keystore without hardware isolation, or cloud workload identity only</td><td>This document</td></tr>
<tr><td>AL2</td><td>Dedicated secure hardware; host software not isolated</td><td>This document</td></tr>
<tr><td>AL3</td><td>TEE that also isolates the agent's code, with measured runtime</td><td>This document</td></tr>
</tbody></table>
</section>
<section anchor="iana-decision-registry"><name>Agent Action Decision Values Registry</name>
<t>Template: Value; Description; Reference. Initial contents:</t>
<table><thead><tr><th>Value</th><th>Description</th><th>Reference</th></tr></thead><tbody>
<tr><td>allowed</td><td>Action signed within the mandate</td><td>This document</td></tr>
<tr><td>step_up_approved</td><td>Human approved the action</td><td>This document</td></tr>
<tr><td>step_up_denied</td><td>Human declined the action</td><td>This document</td></tr>
<tr><td>denied_out_of_scope</td><td>Action outside the mandate</td><td>This document</td></tr>
<tr><td>denied_origin_mismatch</td><td>Origin did not match the credential scope</td><td>This document</td></tr>
<tr><td>denied_revoked</td><td>Credential or mandate revoked</td><td>This document</td></tr>
</tbody></table>
</section>
</section>
</section>
</middle>
<back>
<references><name>References</name>
<references><name>Normative References</name>
<reference anchor="RFC2119" target="https://www.rfc-editor.org/info/rfc2119"><front><title>Key words for use in RFCs to Indicate Requirement Levels</title><author initials="S." surname="Bradner" fullname="S. Bradner"/><date year="1997" month="March"/></front><seriesInfo name="BCP" value="14"/><seriesInfo name="RFC" value="2119"/><seriesInfo name="DOI" value="10.17487/RFC2119"/></reference>
<reference anchor="RFC8174" target="https://www.rfc-editor.org/info/rfc8174"><front><title>Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words</title><author initials="B." surname="Leiba" fullname="B. Leiba"/><date year="2017" month="May"/></front><seriesInfo name="BCP" value="14"/><seriesInfo name="RFC" value="8174"/><seriesInfo name="DOI" value="10.17487/RFC8174"/></reference>
<reference anchor="RFC7515" target="https://www.rfc-editor.org/info/rfc7515"><front><title>JSON Web Signature (JWS)</title><author initials="M." surname="Jones" fullname="M. Jones"/><author initials="J." surname="Bradley" fullname="J. Bradley"/><author initials="N." surname="Sakimura" fullname="N. Sakimura"/><date year="2015" month="May"/></front><seriesInfo name="RFC" value="7515"/><seriesInfo name="DOI" value="10.17487/RFC7515"/></reference>
<reference anchor="RFC7638" target="https://www.rfc-editor.org/info/rfc7638"><front><title>JSON Web Key (JWK) Thumbprint</title><author initials="M." surname="Jones" fullname="M. Jones"/><author initials="N." surname="Sakimura" fullname="N. Sakimura"/><date year="2015" month="September"/></front><seriesInfo name="RFC" value="7638"/><seriesInfo name="DOI" value="10.17487/RFC7638"/></reference>
<reference anchor="RFC9052" target="https://www.rfc-editor.org/info/rfc9052"><front><title>CBOR Object Signing and Encryption (COSE): Structures and Process</title><author initials="J." surname="Schaad" fullname="J. Schaad"/><date year="2022" month="August"/></front><seriesInfo name="STD" value="96"/><seriesInfo name="RFC" value="9052"/><seriesInfo name="DOI" value="10.17487/RFC9052"/></reference>
<reference anchor="RFC8693" target="https://www.rfc-editor.org/info/rfc8693"><front><title>OAuth 2.0 Token Exchange</title><author initials="M." surname="Jones" fullname="M. Jones"/><author initials="A." surname="Nadalin" fullname="A. Nadalin"/><author initials="B." surname="Campbell" fullname="B. Campbell" role="editor"/><author initials="J." surname="Bradley" fullname="J. Bradley"/><author initials="C." surname="Mortimore" fullname="C. Mortimore"/><date year="2020" month="January"/></front><seriesInfo name="RFC" value="8693"/><seriesInfo name="DOI" value="10.17487/RFC8693"/></reference>
<reference anchor="RFC7523" target="https://www.rfc-editor.org/info/rfc7523"><front><title>JSON Web Token (JWT) Profile for OAuth 2.0 Client Authentication and Authorization Grants</title><author initials="M." surname="Jones" fullname="M. Jones"/><author initials="B." surname="Campbell" fullname="B. Campbell"/><author initials="C." surname="Mortimore" fullname="C. Mortimore"/><date year="2015" month="May"/></front><seriesInfo name="RFC" value="7523"/><seriesInfo name="DOI" value="10.17487/RFC7523"/></reference>
<reference anchor="RFC9396" target="https://www.rfc-editor.org/info/rfc9396"><front><title>OAuth 2.0 Rich Authorization Requests</title><author initials="T." surname="Lodderstedt" fullname="T. Lodderstedt"/><author initials="J." surname="Richer" fullname="J. Richer"/><author initials="B." surname="Campbell" fullname="B. Campbell"/><date year="2023" month="May"/></front><seriesInfo name="RFC" value="9396"/><seriesInfo name="DOI" value="10.17487/RFC9396"/></reference>
<reference anchor="RFC9449" target="https://www.rfc-editor.org/info/rfc9449"><front><title>OAuth 2.0 Demonstrating Proof of Possession (DPoP)</title><author initials="D." surname="Fett" fullname="D. Fett"/><author initials="B." surname="Campbell" fullname="B. Campbell"/><author initials="J." surname="Bradley" fullname="J. Bradley"/><author initials="T." surname="Lodderstedt" fullname="T. Lodderstedt"/><author initials="M." surname="Jones" fullname="M. Jones"/><author initials="D." surname="Waite" fullname="D. Waite"/><date year="2023" month="September"/></front><seriesInfo name="RFC" value="9449"/><seriesInfo name="DOI" value="10.17487/RFC9449"/></reference>
<reference anchor="RFC9334" target="https://www.rfc-editor.org/info/rfc9334"><front><title>Remote ATtestation procedureS (RATS) Architecture</title><author initials="H." surname="Birkholz" fullname="H. Birkholz"/><author initials="D." surname="Thaler" fullname="D. Thaler"/><author initials="M." surname="Richardson" fullname="M. Richardson"/><author initials="N." surname="Smith" fullname="N. Smith"/><author initials="W." surname="Pan" fullname="W. Pan"/><date year="2023" month="January"/></front><seriesInfo name="RFC" value="9334"/><seriesInfo name="DOI" value="10.17487/RFC9334"/></reference>
<reference anchor="RFC9711" target="https://www.rfc-editor.org/info/rfc9711"><front><title>The Entity Attestation Token (EAT)</title><author initials="L." surname="Lundblade" fullname="L. Lundblade"/><author initials="G." surname="Mandyam" fullname="G. Mandyam"/><author initials="J." surname="O'Donoghue" fullname="J. O'Donoghue"/><author initials="C." surname="Wallace" fullname="C. Wallace"/><date year="2025" month="April"/></front><seriesInfo name="RFC" value="9711"/><seriesInfo name="DOI" value="10.17487/RFC9711"/></reference>
<reference anchor="WebAuthn" target="https://www.w3.org/TR/2026/REC-webauthn-3-20260825/"><front><title>Web Authentication: An API for accessing Public Key Credentials Level 3</title><author initials="T." surname="Cappalli" fullname="T. Cappalli" role="editor"/><author initials="A." surname="Kumar" fullname="A. Kumar" role="editor"/><author initials="E." surname="Lundberg" fullname="E. Lundberg" role="editor"/><author initials="M." surname="Miller" fullname="M. Miller" role="editor"/><author surname="Pascoe" fullname="Pascoe" role="editor"/><author initials="N." surname="Satragno" fullname="N. Satragno" role="editor"/><date year="2026" month="August" day="25"/></front><seriesInfo name="W3C" value="REC-webauthn-3-20260825"/><annotation>Latest version available at <eref target="https://www.w3.org/TR/webauthn-3/"/>.</annotation></reference>
</references>
<references><name>Informative References</name>
<reference anchor="RFC6838" target="https://www.rfc-editor.org/info/rfc6838"><front><title>Media Type Specifications and Registration Procedures</title><author initials="N." surname="Freed" fullname="N. Freed"/><author initials="J." surname="Klensin" fullname="J. Klensin"/><author initials="T." surname="Hansen" fullname="T. Hansen"/><date year="2013" month="January"/></front><seriesInfo name="BCP" value="13"/><seriesInfo name="RFC" value="6838"/><seriesInfo name="DOI" value="10.17487/RFC6838"/></reference>
<reference anchor="RFC7942" target="https://www.rfc-editor.org/info/rfc7942"><front><title>Improving Awareness of Running Code: The Implementation Status Section</title><author initials="Y." surname="Sheffer" fullname="Y. Sheffer"/><author initials="A." surname="Farrel" fullname="A. Farrel"/><date year="2016" month="July"/></front><seriesInfo name="BCP" value="205"/><seriesInfo name="RFC" value="7942"/><seriesInfo name="DOI" value="10.17487/RFC7942"/></reference>
<reference anchor="RFC8126" target="https://www.rfc-editor.org/info/rfc8126"><front><title>Guidelines for Writing an IANA Considerations Section in RFCs</title><author initials="M." surname="Cotton" fullname="M. Cotton"/><author initials="B." surname="Leiba" fullname="B. Leiba"/><author initials="T." surname="Narten" fullname="T. Narten"/><date year="2017" month="June"/></front><seriesInfo name="BCP" value="26"/><seriesInfo name="RFC" value="8126"/><seriesInfo name="DOI" value="10.17487/RFC8126"/></reference>
<reference anchor="RFC9943" target="https://www.rfc-editor.org/info/rfc9943"><front><title>An Architecture for Trustworthy and Transparent Digital Supply Chains</title><author initials="H." surname="Birkholz" fullname="H. Birkholz"/><author initials="A." surname="Delignat-Lavaud" fullname="A. Delignat-Lavaud"/><author initials="C." surname="Fournet" fullname="C. Fournet"/><author initials="Y." surname="Deshpande" fullname="Y. Deshpande"/><author initials="S." surname="Lasker" fullname="S. Lasker"/><date year="2026" month="April"/></front><seriesInfo name="RFC" value="9943"/><seriesInfo name="DOI" value="10.17487/RFC9943"/></reference>
<reference anchor="ID-JAG" target="https://datatracker.ietf.org/doc/html/draft-ietf-oauth-identity-assertion-authz-grant-04"><front><title>Identity Assertion JWT Authorization Grant</title><author initials="A." surname="Parecki" fullname="A. Parecki"/><author initials="K." surname="McGuinness" fullname="K. McGuinness"/><author initials="B." surname="Campbell" fullname="B. Campbell"/><date year="2026" month="May"/></front><seriesInfo name="Internet-Draft" value="draft-ietf-oauth-identity-assertion-authz-grant-04"/></reference>
<reference anchor="AIMS" target="https://datatracker.ietf.org/doc/html/draft-ietf-wimse-aims-00"><front><title>AI Identity Management System</title><author initials="P." surname="Kasselman" fullname="P. Kasselman"/><author initials="J." surname="Lombardo" fullname="J. Lombardo"/><author initials="Y." surname="Rosomakho" fullname="Y. Rosomakho"/><author initials="B." surname="Campbell" fullname="B. Campbell"/><author initials="N." surname="Steele" fullname="N. Steele"/><author initials="A." surname="Parecki" fullname="A. Parecki"/><date year="2026" month="September"/></front><seriesInfo name="Internet-Draft" value="draft-ietf-wimse-aims-00"/></reference>
<reference anchor="SSF" target="https://openid.net/specs/openid-sharedsignals-framework-1_0-final.html"><front><title>OpenID Shared Signals Framework Specification 1.0</title><author initials="A." surname="Tulshibagwale" fullname="A. Tulshibagwale"/><author initials="T." surname="Cappalli" fullname="T. Cappalli"/><author initials="M." surname="Scurtescu" fullname="M. Scurtescu"/><author initials="A." surname="Backman" fullname="A. Backman"/><author initials="J." surname="Bradley" fullname="J. Bradley"/><author initials="S." surname="Miel" fullname="S. Miel"/><date year="2025" month="August" day="29"/></front></reference>
<reference anchor="MCP-AUTH" target="https://modelcontextprotocol.io/"><front><title>Model Context Protocol: Authorization</title><author><organization>Model Context Protocol</organization></author><date year="2026"/></front></reference>
<reference anchor="OCSF" target="https://schema.ocsf.io/"><front><title>Open Cybersecurity Schema Framework</title><author><organization>OCSF Project</organization></author><date year="2026"/></front></reference>
<reference anchor="IANA.MEDIA" target="https://www.iana.org/assignments/media-types"><front><title>Media Types</title><author><organization>IANA</organization></author><date/></front></reference>
<reference anchor="IANA.OAUTH" target="https://www.iana.org/assignments/oauth-parameters"><front><title>OAuth Parameters</title><author><organization>IANA</organization></author><date/></front></reference>
<reference anchor="IANA.JWT" target="https://www.iana.org/assignments/jwt"><front><title>JSON Web Token (JWT)</title><author><organization>IANA</organization></author><date/></front></reference>
<reference anchor="INVARIANT" target="https://invariantlabs.ai/blog/mcp-github-vulnerability"><front><title>GitHub MCP Exploited: Accessing private repositories via MCP</title><author><organization>Invariant Labs</organization></author><date year="2025" month="May"/></front></reference>
<reference anchor="CSA-CI" target="https://labs.cloudsecurityalliance.org/wp-content/uploads/2026/08/CSA_research_note_ai_coding_agent_ci_prompt_injection_20260807-csa-styled.pdf"><front><title>Research Note: AI Coding Agent CI Prompt Injection</title><author><organization>Cloud Security Alliance</organization></author><date year="2026" month="August" day="7"/></front></reference>
<reference anchor="VB-EXPLOITS" target="https://venturebeat.com/security/six-exploits-broke-ai-coding-agents-iam-never-saw-them"><front><title>Six exploits broke AI coding agents. IAM never saw them</title><author><organization>VentureBeat</organization></author><date year="2026"/></front></reference>
<reference anchor="NX-LEAK" target="https://blog.gitguardian.com/the-nx-s1ngularity-attack-inside-the-credential-leak/"><front><title>The Nx s1ngularity Attack: Inside the Credential Leak</title><author><organization>GitGuardian</organization></author><date year="2025"/></front></reference>
<reference anchor="SHAI-HULUD" target="https://www.elastic.co/security-labs/shai-hulud-chaindrop-npm-supply-chain"><front><title>Shai-Hulud Strikes Again: CHAINDROP Worm Hits 400+ npm Packages</title><author><organization>Elastic Security Labs</organization></author><date year="2026"/></front></reference>
<reference anchor="ASTRIX" target="https://astrix.security/learn/news/astrix-researchers-uncover-credential-risk-in-the-majority-of-mcp-servers/"><front><title>Astrix Researchers Uncover Credential Risk in the Majority of MCP Servers</title><author><organization>Astrix Security</organization></author><date year="2025"/></front></reference>
<reference anchor="GG-SPRAWL" target="https://www.gitguardian.com/state-of-secrets-sprawl-report-2026"><front><title>The State of Secrets Sprawl 2026</title><author><organization>GitGuardian</organization></author><date year="2026"/></front></reference>
</references>
</references>
<section anchor="standardization-path" removeInRFC="true"><name>Standardization Path</name>
<t>The specification may be split into smaller documents, each sent to the group that owns that layer.</t>
<table><thead><tr><th>Document</th><th>Venue</th><th>Contents</th></tr></thead><tbody><tr><td>EAT profile for agent keys</td><td>IETF RATS</td><td>Vendor evidence to AL0-AL3 mapping</td></tr><tr><td>Agent assertion and mandate profile</td><td>IETF OAuth, with WIMSE</td><td>AgentData, mandate, token exchange and ID-JAG use</td></tr><tr><td>Delegated approval extension</td><td>W3C WebAuthn WG, FIDO Alliance</td><td>Passkey signing of mandates and per-action approvals</td></tr><tr><td>Agent action-record profile</td><td>IETF SCITT, plus OCSF</td><td>Records, chain, anchoring and OCSF mapping</td></tr><tr><td>MCP binding</td><td>MCP authorization working groups</td><td>DPoP and workload identity usage with the AA</td></tr></tbody></table>
</section>
<section anchor="related-work"><name>Related Work</name>
<t>The AI Identity Management System <xref target="AIMS"/> composes workload identity, OAuth and shared signals into a model for agent authentication and authorization, treating agents as workloads; this document adds hardware-bound agent keys, runtime-enforced origin binding and human approval with passkeys. The Identity Assertion JWT Authorization Grant <xref target="ID-JAG"/> and MCP authorization <xref target="MCP-AUTH"/> let identity providers control which applications an agent may reach; this document supplies them with stronger evidence about the agent rather than replacing them. Transparency services <xref target="RFC9943"/> provide the anchoring used for action records.</t>
</section>
<section anchor="acknowledgments"><name>Acknowledgments</name>
<t>This draft builds on the WebAuthn, OAuth, RATS, WIMSE and SCITT work of the W3C, the FIDO Alliance and the IETF.</t>
</section>
</back>
</rfc>
