Revocation Exposure
The operational risk that remains after authority is revoked but before every covered execution path stops accepting it: completion time × accepted action rate × irreversible impact.
Revocation Exposure is the operational risk accumulated after a central revocation decision but before every covered execution path has stopped honoring the authority. It is estimated as Revocation Completion Time multiplied by the accepted action rate and the average irreversible impact per action. When impact cannot be normalized honestly, report the simpler and still useful quantity: actions at risk.
The metric corrects a blind spot in token-lifetime discussions. A five-minute stale window is not one security level: it may expose one human action or thousands of autonomous tool calls. Sender-constraining with DPoP or mTLS reduces credential portability, while expiry, introspection, deny state and continuous-access events reduce stale authority. Revocation Exposure measures the residual outcome after those controls interact with the system's actual execution rate.
To measure it, define the covered gateways, APIs, regions, connections, workers and descendants; trigger revocation while the credential is active; continue exercising every path until each denies it; then report p50, p95, p99 and maximum Revocation Completion Time by operation class. Multiply the relevant tail by accepted actions per second. For AI agents, this turns revocation from an endpoint checkbox into a falsifiable emergency-brake SLO expressed in both seconds and irreversible actions at risk.
Frequently asked questions
What is Revocation Exposure?
The risk accumulated between a central revocation decision and complete enforcement. It is estimated as Revocation Completion Time × accepted action rate × average irreversible impact; when impact cannot be normalized, report actions at risk.
Why is token lifetime alone insufficient for AI agents?
Because the same time window contains radically different amounts of execution. A human may perform one meaningful action in a minute, while an autonomous agent can perform hundreds or thousands, so the security bound must combine time with accepted action rate.
How do you measure Revocation Exposure?
Exercise a credential through every covered region and enforcement path, trigger revocation, record the last accepted action on each path, calculate tail Revocation Completion Time, then multiply it by the accepted action rate and, where defensible, irreversible impact.
Source analyses
Related terms
Term defined and maintained by Michał Piszczek — CTO of Archdesk, author of the Joule Wars and Proof-Adjusted Autonomy frameworks. Quote freely with attribution to piszczek.pl/glossary (CC BY 4.0).