Agentic JWT
AJWT · agentic JWT
A proposed JWT extension that binds an AI agent's specific intent — the task it was delegated — into the access token, so services can verify not just who an agent acts for but what it was authorized to do. An arXiv preprint and IETF individual draft, not a ratified standard.
In our briefings
The claim each of the newest briefings on Agentic JWT would have you leave with.
Mattson, SecureAuth's chief executive, argues that a model cannot be secured the way tested software is: it is non-deterministic, and a request phrased another way, or in another language, gets around whatever it was told not to do. His advice is to "just assume that it could do anything any time" and put the guardrails around the model, so that a financial transaction or a database change the agent attempts is approved and recorded. He cites Anthropic as saying no guardrail is adequate. The statement we take him to mean, Anthropic's of 12 June, which he recalls as "three weeks ago, I believe", says the company suspects perfect jailbreak resistance "is not currently possible for any model provider", while defending its own safeguards as strong. Briefings we have published on other sources move the control off the model too: a Silverfort guest calls guardrails on a probabilistic system guidance rather than control, ClearVector's John Laliberte compares them to bowling bumpers, and a zero-trust framework paper for agents adds a session-enforcement layer.
When the model cannot be trusted, control what the agent can reach, checked on every actionBI-FIND-2026-0151
Revision -02 adds a section on what a server holding the credentials cannot promise. Compromise of the Delegation Server (DS), its operator or the path that releases the keys protecting stored credentials lets an attacker use those credentials with no Delegation Token at all, and Sweeney writes that "Keeping credentials away from Agents does not remove this exposure; it concentrates it at the DS." The server's signed receipts prove it authorized an action under its own records, but, in his words, "it does not prove that the Subject, rather than the operator, approved the grant". Approval signed with a key the user holds is left to a later revision. The revocation section now states its limits too. The one-second and five-second targets stand, yet "revoked_count reports invalidation, not closure": a token the server exchanged for direct use at a resource server, or issued under the offline profile, can be honored by an enforcement point off the server's path until it expires, at most an hour after issue, unless that point checks authority or status. Synchronous cascading revocation holds within one Delegation Server; enforcing it across servers or domains is out of scope.
Sweeney's delegation draft concedes its server is trusted with users' credentials, and drops an interop claimBI-FIND-2026-1939
The draft's sharpest rule reads "The Large Language Model MUST NOT have access to an agent's credentials or to credentials that may be needed to access tools and services", which it says keeps prompt injection from getting the model to use, expose or disclose them. Around it sit the identity rules. An agent MUST be assigned exactly one WIMSE identifier and MUST hold credentials cryptographically bound to it; those credentials SHOULD be short-lived and MUST carry an explicit expiry, and static API keys are "an antipattern for agent identity". The draft names WIT, X.509 WIC and SPIFFE SVIDs as formats and leaves the choice to the deployment. Our briefing on Sierra's cloud agent platform found a similar boundary built in: a proxy approves each call and swaps in the credential, so the agent never holds it.
Six companies' agent-auth draft, since adopted by the WIMSE group, keeps credentials out of the model's reachBI-FIND-2026-0179
›Cited in 11 briefings
Every briefing that discusses this term, newest first.
- When the model cannot be trusted, control what the agent can reach, checked on every action
Podcast · · Tech Talks Daily · AI Agent Security: Why Identity and Access Control Matter More Than Guardrails · Aug 11, 2026
- Sweeney's delegation draft concedes its server is trusted with users' credentials, and drops an interop claim
Draft · · IETF Datatracker · Kieran Sweeney · Oct 13, 2026
- Six companies' agent-auth draft, since adopted by the WIMSE group, keeps credentials out of the model's reach
Draft · · IETF Internet-Draft (individual submission, since adopted by the WIMSE working group) · Pieter Kasselman et al. · Jul 6, 2026
- Cisco's SD-Card puts an A2A Agent Card in a selective-disclosure token, and leaves its registry out of scope
Draft · · IETF Datatracker · Suhas Nandakumar et al. · Mar 2, 2026
- The 'intent' in the Agentic JWT draft is a workflow step, and the agent reports its own fingerprint
Draft · · IETF Datatracker · Abhishek Goswami · Jul 1, 2026
- An Imperial and Oxford protocol would give AI agents one approved use of a secret, never the secret itself
Paper · · arXiv · Xiaohang Yu et al. · Apr 27, 2026
- A proposed record of who authorized an AI agent's actions cannot yet prove which agent took them
Paper · · arXiv · Asiri Dalugoda, Helixar · Apr 6, 2026
- An agent-credential draft carries a hash of the user's instruction that its own verification steps never check
Draft · · IETF Datatracker · Chudah Yakung, Attest · Mar 26, 2026
- A January design paper argued AP2 and x402 move money before anyone checks the work, and proposed escrow
Paper · · arXiv · Mehul Goenka et al. · Jan 30, 2026
- Agentic JWT binds an agent's token to its task and its own code, but it is one author's draft, not a standard
Paper · · arXiv (cs.CR preprint 2509.13597) · Abhishek Goswami · Sep 16, 2025
- In GNAP (RFC 9635), a client is known by its key, and tokens stay key-bound unless the server flags them bearer
Article · · RFC Editor · Justin Richer et al. · Oct 1, 2024
Related terms
What appears alongside Agentic JWT in the same briefings.
Model Context ProtocolLangGraphOAuth for AI Agents on Behalf of UsersGNAPDelegate SD-JWTCrewAIUCANDefakto