Skip to content
IAM Insights

Explainers · Glossary

Capability tokens

Biscuit tokens · object-capability · OCap

Tokens such as Biscuits and Macaroons that carry their own limits, so the holder can hand on a narrower copy offline without asking the issuer.

In our briefings

The claim each of the newest briefings on Capability tokens would have you leave with.

  1. Burls's credential records what an agent was issued and what it declared, not what it may do. The leaf field is named authorized_scopes, yet its capability tokens are self-reported by the requesting entity and recorded by the CA, and the draft defines no vocabulary for them and no mechanism by which they constrain behaviour. It says a relying party "MUST NOT treat the presence of a scope token as evidence that the agent operated within it." Burls calls this issuance-time boundary the most important consideration in the document: a credential that verifies may be held by an agent whose behaviour has since changed, and assurance about request time needs a mechanism outside MTAC.

    A post-quantum agent certificate draft warns that its own declared scopes are not proof of behaviourBI-FIND-2026-1138

  2. In the Accenture design an agent's identity is a key pair shown through a DID document and domain control, and the user grants a two-hour master capability that each agent narrows before passing on (USD 5,000 to the trip planner, USD 1,500 hold-only to the airline agent). Keys and credentials live in a wallet process outside the language model's context, reached only through typed calls. But the user's own wallet issues the credentials it later presents, and the paper names the user the "trust anchor": nobody outside proofs the person. The service checks attenuation by a rule the paper calls domain-specific and never writes down, and revocation is left as three options (five-minute tokens, a walk up the parent chain, or a cryptographic accumulator) with none chosen.

    In Accenture's agent-delegation design, no outside party checks the user, and revocation is left openBI-FIND-2026-0667

  3. The OpenID Foundation's own community group draws the line on what today's identity standards do for AI agents: they 'can work for synchronous agents using multiple tools across a single trust domain, but not asynchronous or multi-domain contexts.' In plain terms, OAuth 2.1 handles an agent that works inside one organization while its user is present. It falls short when the agent runs unattended, crosses into another organization's systems, or acts for several people at once. Everything vendors now sell as agent identity for the simple case rests on standards that already existed; the hard cases have drafts and proposals.

    OAuth handles an agent in one company, working while you watch. The OpenID paper says that is where it stops.BI-FIND-2025-0054

Cited in 5 briefings

Every briefing that discusses this term, newest first.

  1. A post-quantum agent certificate draft warns that its own declared scopes are not proof of behaviour

    Draft · · IETF Datatracker · Reuben Burls · Jul 21, 2026

  2. In Accenture's agent-delegation design, no outside party checks the user, and revocation is left open

    Paper · · arXiv · Srikumar K. Subramanian et al. · Sep 25, 2026

  3. OAuth handles an agent in one company, working while you watch. The OpenID paper says that is where it stops.

    Article · · OpenID Foundation · Tobin South et al. · Oct 2025

  4. A proposed standard for agent authority advises against delegating it across company lines by default

    Paper · · arXiv · Partha Madhira · May 12, 2026

  5. Google DeepMind wants agent authority to narrow at every hand-off, and misreads MCP's scopes on the way

    Paper · · arXiv · Nenad Tomašev et al. · Feb 12, 2026