NIST’s Ten
Tenets for
Agent Identity.
Revoked is not stopped. Two standards bodies now describe the same agent identity design, and the off switch is the part nobody has proven.
On October 27, NIST published ten design tenets for managing the identity of AI agents. On November 2, an IETF working group revised its reference model for how agents authenticate and get authorized. Read side by side, the two documents describe one design.
An identity of the agent’s own. Authority that narrows at every hand-off. Credentials that are short-lived and bound to a key the model never holds. Policy decided outside the model’s reasoning. Evidence the agent cannot reach. If you are building or buying agent identity, that is now the reference picture. It is worth knowing exactly what each document says, what each one does not, and what two new papers found when they tested the hard parts.
What NIST Said
The post is “Design Tenets for Agent Identity Management Systems”, by Harold Booth, Bill Fisher and Ryan Galluzzo of NIST and Tom Clancy of MITRE, the second in a NIST blog series on agentic identity. It proposes, in its own words, “an initial set of ten design tenets”, drawn from public comments to the National Cybersecurity Center of Excellence.
- 01 Distinct Agent Identification. Unique, cryptographically verifiable identification of agents, distinguishable from human users and service accounts.
- 02 Agent Linking, Delegation & Attenuation. Agents stay linked to whoever started the task, and “scopes attenuate so that downstream entities can never claim more authority than the parent”.
- 03 Least & Zero Standing Privilege. “Agents operate with zero standing privilege, holding no persistent rights when idle.”
- 04 Short-Lived, Dynamic & Revocable Authorization. “All access grants are Just-In-Time (JIT), time-bounded and audience-specific.” No long-lived bearer tokens or static secrets.
- 05 Human-in-the-Loop Triggers. Risk thresholds that, when reached, bring a person into the decision.
- 06 Mandates and Constraints. The task originator can issue “cryptographically signed mandate statements” that define the intended outcome and its limits.
- 07 Architectural Separation of Policy from Reasoning. What the agent may do is decided outside the logic that decides what it tries to do.
- 08 Deterministic Access Decisions. Deterministic capabilities “as the primary mechanism for making access decisions”. Probabilistic signals can inform, not decide.
- 09 Continuous Monitoring. Agent-specific metrics, thresholds and policies, because agents run at a scale and speed people cannot.
- 10 Immutable Evidence. An “unbroken, tamper-evident chain of custody and append-only body of evidence”, logged independently of the agent.
Three things to keep in view. It is a blog post, not a Special Publication. NIST says plainly that “not all of these design tenets will be immediately fulfilled by the current state of standards and technology”, and that they “will be the guideposts for our agentic demonstrations in the NCCoE”. And it takes no position on certificates versus tokens. The word certificate does not appear.
What the IETF Said
The IETF document is draft-ietf-wimse-aims-01, “AI Identity Management System”, from the Workload Identity in Multi System Environments working group, dated November 2. Its six authors work at Defakto Security, AWS, Zscaler, Ping Identity, OpenAI and Okta. It is a reference model, not a protocol: “Rather than defining new protocols”, it profiles existing ones, WIMSE and the OAuth 2.0 family. It is an Informational draft at an early stage, with no document shepherd or area director assigned yet, and it expires on May 6, 2027 unless revised.
Six of its lines carry most of the weight. “An agent participating in this framework MUST be assigned exactly one WIMSE identifier, which MAY be a SPIFFE ID.” “Agent credentials SHOULD be short-lived to minimize the risk of credential theft”, and static API keys “are an antipattern for agent identity”. “The Large Language Model MUST NOT have access to an agent’s credentials”. When one agent hands work to another through token exchange, “The scope of access conveyed by each token in this sequence of exchanges can never exceed that of the token presented to obtain it.” “Audit records MUST be tamper-evident”.
And on switching an agent off: “At a minimum, deprovisioning stops renewal so that outstanding short-lived credentials expire and are not replaced.” Where that is too slow, the issuer signals the change to relying parties and to the issuers of any derived credentials, through eventing such as the OpenID Shared Signals Framework. “Recipients of such signals MUST ensure that revoked or downgraded authorization is enforced without undue delay.” Cached tokens that are no longer valid “MUST NOT continue to be used” once a revocation notice arrives.
Where They Agree
The two documents were written by different people for different purposes, a federal lab setting guideposts for demonstrations and a working group profiling existing protocols. Agreement at this level is new. A year ago the honest answer to “what should agent identity look like” was a dozen vendor diagrams. Now there is a reference picture with two signatures on it.
The Part Nobody Has Proven
Two papers from the last eight weeks tested the two hardest rows of that table: delegation that narrows, and switching off. Neither result is comfortable.
Delegation. Gabriele Donno, Kashif Mehmood and Katina Kralevska of NTNU in Trondheim built a layered design on the Agent2Agent protocol and tested it in a synthetic four-party healthcare reimbursement workflow: per-hop request signing with Ed25519 keys, a separate piece of authorization evidence that travels with the request, and replay protection. It held up well. Of 15 adversarial conditions, 14 were rejected, at a median of 5.11 ms of security processing per request and a 3.87 kB payload.
The two failures are the lesson. First, authority did not narrow from hop to hop: a test that widened the scope between two agents was accepted, because the path check compares structure, not successive scopes. Second, the authorization evidence was protected by a key shared among the participants. An insider holding that key, the hospital agent in the test, changed the authorized maximum from 120 to 9,000 euros, recomputed the evidence and passed every check. In the authors’ words, “every holder of K_A can cryptographically generate authorization evidence that is accepted as belonging to the same trust domain.” Evidence under a shared key proves membership of a group. It does not prove what the originator authorized. That is NIST’s tenet 02 and the IETF’s “can never exceed” rule with a failure mode attached: delegation has to be signed by its origin and narrowed at each hop, or it is neither. The caveats: the workflow is synthetic, local and of fixed depth, and key rotation, revocation and PKI are out of scope.
Switching off. Genliang Zhu and Chu Wang of Accentrust in Vancouver start from the three things an operator can see after pulling the switch: the cancellation was acknowledged, the token was revoked, the process exited. Their claim is that none of those, and not all three together, prove the agent has stopped: “Declaring quiescence from the three local observations is false.” A queued message, a delayed callback, a retry record, a child process or a credential the agent delegated can each carry authority past the point where the first credential died. “Revoking the initiating credential therefore closes one carrier, not necessarily the entire root cut.”
Their answer is a proof rather than a timer. Cut the agent’s root authority on a ledger, put a fence at every place an effect can land, collect signed receipts from each, and compose them into one of three verdicts: Quiescent, Not Quiescent or Indeterminate. Only Quiescent releases. Neither of the others becomes success through a timeout, exhausted retries, a majority vote or the agent’s own claim that it has stopped, and an endpoint that cannot be fenced stays Indeterminate rather than disappearing. On a provider-free test suite their runner matched 17 of 17 expected outcomes, and a separately written checker verified 17 of 17 traces and rejected 44 of 44 tampered ones. The caveats: nothing ran against a real cloud provider, no latency was measured, the safety argument rests on eleven stated assumptions, and any endpoint you cannot fence comes out Indeterminate. The point survives the caveats.
Why This Matters for the Off Switch
In FN-14 we showed that a kill switch which asks the agent to stop works less than half the time. In FN-16 we showed why the banks had only one switch, a wall, because an agent with a borrowed password cannot be revoked on its own. The switch both standards bodies describe is the right one: revoke the agent’s own credential. NIST’s reasoning for tenet 04 is that “an agent operating under a valid token cannot be interrupted”, so keep tokens short and make them revocable. The IETF adds the plumbing: stop renewal, send the signal, enforce it without undue delay, drop cached tokens.
Every one of those steps is about the next action. Zhu and Wang are about the actions already in flight. That is the gap between revoked and stopped, and no standard, product or paper has closed it yet. The next thing to ask for from anyone selling an off switch is not a smaller number of seconds. It is a receipt.
Where AgentID Stands
Here is AgentID against NIST’s ten, as it ships today. We would rather you read this than a brochure.
On the hardest row, we will say exactly what we can prove. Revoking an agent’s identity in AgentID means the next action that passes through its gateway is refused. It does not prove that work the agent already queued elsewhere has stopped, and it does not cover actions that never pass through the gateway. We do not publish a stop-time figure, and we will not until we can show a verdict an auditor can check, which is what Zhu and Wang describe.
Three Questions for Anyone Selling an Off Switch
- 1. Revoke what? A credential, a session, a process? Each one closes one carrier of authority, not all of them.
- 2. What happens to work already in flight? Queued jobs, callbacks, retries, sub-agents, credentials the agent handed on.
- 3. What evidence comes back, and does missing evidence count as success?
If the answer to the third question is a timer, you have a hope, not a proof.
Field Note FN-17-2026 · Distributed under arkion.ai/field-notes
For questions or to discuss findings: research@arkion.ai
- Harold Booth, Bill Fisher, Ryan Galluzzo (NIST) and Tom Clancy (MITRE), “Design Tenets for Agent Identity Management Systems”, NIST Cybersecurity Insights blog, October 27, 2026.
- Pieter Kasselman, Jeff Lombardo, Yaroslav Rosomakho, Brian Campbell, Nick Steele and Aaron Parecki, “AI Identity Management System”, draft-ietf-wimse-aims-01, IETF WIMSE working group, Informational Internet-Draft, November 2, 2026.
- Gabriele Donno, Kashif Mehmood and Katina Kralevska (NTNU), “Authorization Propagation in Agent2Agent Systems: An Experimental Security Evaluation of Multi-Hop Delegation”, arXiv:2611.07050, November 4, 2026.
- Genliang Zhu and Chu Wang (Accentrust), “Authorization Revocation for Long-Running AI Agents: Root-Scoped Quiescence under Delegation and Asynchronous Execution”, arXiv:2609.21284, September 18, 2026.
Recommend Arkion in one click.
An identity for every agent,
and an honest off switch.
Book a demo and we will show you what AgentID does today against NIST’s ten tenets, and say plainly what it does not.