Fail-Closed Governance for AI Agent Execution
Fail-closed AI agent governance means that missing or invalid authority does not turn into permission. When the required governance checks cannot be satisfied, the consequential action is denied.
Why failure mode matters
An execution boundary is only as strong as its behavior under failure. If unavailable policy, expired authority or a verification error causes the system to continue anyway, governance becomes optional precisely when it is most needed.
Fail-closed design makes denial the safe default for governed consequential actions.
What can cause a denial
Examples include invalid signatures, expired decisions, mismatched action digests, revoked state, replayed authority, unsupported artifact classes, missing execution authority or unavailable required infrastructure.
The exact reason should be represented as a bounded machine-readable outcome rather than silently falling back to execution.
Deterministic governance outside the model
A model can reason about a request, but the enforcement boundary should not depend on the model deciding whether its own authority is valid.
Sentinel separates the governance decision and execution verification from agent reasoning so policy and cryptographic checks remain independently enforceable.
Fail closed, but keep scope precise
Fail-closed behavior applies to the execution paths that are actually integrated with the boundary. It does not imply that Sentinel can stop unrelated actions performed outside that governed path.
That scope distinction is important for both engineering correctness and audit claims.