Token replay
token theft · stolen token replay · refresh token theft
Using a stolen access, refresh or session token from the attacker's own machine to act as the user without re-authenticating.
In our briefings
The claim each of the newest briefings on Token replay would have you leave with.
Niyikiza's tokens let one AI agent narrow another's authority down to the values a tool's arguments may take. A root token from the authorization server maps tool names to constraints on each argument (exact, range, one_of, subset and five more core types), and a holder can derive a child offline with fewer tools or tighter constraints, sign it with its own key and pass it on. Argument matching stays closed by default in revision -02: the enforcement point refuses an argument the token does not name unless the issuer adds a "*" entry, which the draft says issuers should not use for tools with side effects. The enforcement point checks the whole chain with only the root issuer's trust anchor and no network call. If it skips or botches the narrowing check, it "provides no blast radius containment", in the draft's words.
Niyikiza's agent tokens narrow to a tool's arguments offline; revision -02 breaks -01 implementationsBI-FIND-2026-0532
In Chu and his Huawei co-authors' draft, conveying a subject's keys grants nothing: "The subject_keys parameter does not grant access." A resource server that checks evidence signed by the token's subject must work out what to allow as "Token Authority INTERSECT Evidence Authority INTERSECT Resource Server Local Policy", and act only when all three permit the whole operation. The signed evidence must not add a scope, resource, action, amount, recipient, time period or delegation right that the token lacks, and the token does not override a narrower limit in the evidence. Where the two are written in different vocabularies, the mapping must be deterministic and fail closed; if it is ambiguous or unsupported, the request is rejected.
In Huawei's reworked OAuth draft, evidence signed by the token's subject can narrow access but never widen itBI-FIND-2026-1154
NIST's final guidance on protecting identity tokens, published with CISA on 15 September 2026, turns token hygiene into numbers. Access and identity tokens should be valid for no more than an hour; signing keys for high-impact systems should be active no more than 90 days and others less than a year; and from moderate impact up, signing keys must sit in hardware-backed or otherwise isolated storage, which a cloud key service satisfies. Keys that sign federal tokens must not sign anything else, workload identities must use tightly scoped short-lived tokens, and tokens must never be written to logs. Binding tokens to the client's key is recommended only for workloads, and the report concedes that stateless tokens cannot always be revoked before they expire. All of it is voluntary.
NIST and CISA advise one-hour tokens, 90-day keys and hardware key storage; AI agents get one paragraphBI-FIND-2026-0700
›Cited in 11 briefings
Every briefing that discusses this term, newest first.
- Niyikiza's agent tokens narrow to a tool's arguments offline; revision -02 breaks -01 implementations
Draft · · IETF Datatracker · Niki Aimable Niyikiza · Jun 15, 2026
- In Huawei's reworked OAuth draft, evidence signed by the token's subject can narrow access but never widen it
Draft · · IETF Datatracker · Cheng-Kang Chu et al. · Sep 1, 2026
- NIST and CISA advise one-hour tokens, 90-day keys and hardware key storage; AI agents get one paragraph
Article · · NIST (CSRC) · Ryan Galluzzo et al. · Sep 15, 2026
- MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extension
Article · · Model Context Protocol project · Jul 28, 2026
- The browser OAuth draft favors a backend that keeps tokens out of JavaScript
Draft · · IETF Datatracker · Aaron Parecki et al. · Jul 6, 2026
- OAuth 2.1's October draft bans plain PKCE and adds browser-app rules; 2.1-only clients can fail on older servers
Draft · · IETF Datatracker · Dick Hardt et al. · Mar 2, 2026
- A poisoned MCP tool got an agent to copy the client's config file without asking; deleting a file did ask
Paper · · arXiv · Yongjian Guo et al. · Aug 18, 2025
- A survey of MCP and A2A attacks pairs each identity threat with a fix, and tests none of them
Paper · · arXiv · Mohamed Amine Ferrag et al. · Jun 29, 2025
- MCP Guardian wraps tool calls in token and regex checks, leaving poisoned servers and agent IDs open
Paper · · arXiv · Sonu Kumar et al. · Apr 17, 2025
- In 2015, OpenID Connect discovery let an attacker's server choose where a client sent the user's code
Paper · · arXiv · Christian Mainka et al. · Aug 18, 2015
- RFC 7009 requires refresh-token revocation but only recommends it for access tokens, which may outlive a revoke
Article · · RFC Editor · Torsten Lodderstedt et al. · Aug 1, 2013