Skip to content
IAM Insights

Explainers · Glossary

TBAC

Task-Based Access Control

An access-control model that scopes an AI agent's permissions to the specific task it has been asked to perform, rather than to a persistent role, attribute set, or static list of OAuth scopes.

In our briefings

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

  1. Shi, Du, Guan and four co-authors start from the view that an agent's beliefs "will always be corruptible" and its intent cannot be verified, so they argue security has to be built at the permission check, the last stage before an action runs. Today's checks, they write, are "largely belief-blind": dynamic access control reacts to changes in context but ignores where the beliefs behind a request came from. Their own tally of the defense literature (Figure 6) puts most papers at the input stage, which they call brittle: neutralization 13 and sanitizing 11, against static least privilege 3, dynamic authorization 3 and taint tracking 4 at the permission stage. Their design labels every input with its source and trust level, makes each plan cite the inputs it relies on, and refuses a high-risk action whose plan cites a low-trust input. Inputs with missing or unverifiable labels count as low trust. Theorem 1 says the rule holds under those conditions, with only a proof sketch and no implementation, and the limitations section says it rests on strong assumptions.

    Refuse the risky action when its plan relies on untrusted input, a Beihang survey of agent attacks proposesBI-FIND-2025-0250

Cited in 1 briefing

Every briefing that discusses this term, newest first.

  1. Refuse the risky action when its plan relies on untrusted input, a Beihang survey of agent attacks proposes

    Paper · · arXiv · Guanquan Shi et al. · Dec 7, 2025