What is reversibility of agentic actions?
Reversibility of agentic actions is the principle that any action an AI agent takes should have a known, tested and fast way to undo it. If an agent isolates the wrong laptop, someone should be able to release it in seconds, and the agent itself should have recorded how.
Reversibility is one of the main controls that make autonomous response acceptable, because it limits the cost of the mistakes every system eventually makes.
Which security actions are reversible?
- Isolate a device. Reversibility: Fully reversible. Notes: Release isolation from the EDR console.
- Quarantine an email. Reversibility: Fully reversible. Notes: Restore to inboxes.
- Block a domain or IP. Reversibility: Fully reversible. Notes: Remove the block; check for side effects while it was in place.
- Disable an account. Reversibility: Mostly reversible. Notes: Re-enable; sessions and some integrations may need re-establishing.
- Force a password reset. Reversibility: Partly reversible. Notes: The old password cannot be restored, but the user can regain access.
- Delete files or wipe a device. Reversibility: Not reversible. Notes: Keep out of automation, or behind dual approval.
How do you make agentic actions reversible?
- Record the undo at the time of the action. The agent writes down exactly how to reverse what it did, including identifiers.
- Prefer containment to destruction. Isolate rather than wipe, disable rather than delete, quarantine rather than purge.
- Test the undo. A reversal path that has never been exercised is a guess.
- Snapshot state where possible, such as group memberships before removing a user from them.
- Gate the irreversible. Anything that cannot be undone needs approval, and often two approvals. See gated approval.
Why does reversibility matter?
Reversibility matters because it changes the risk calculation for automation. An action that is fast and fully reversible can often run without waiting for approval, since the worst case is a short disruption. An action that cannot be undone needs much stronger evidence and a person's sign-off. Teams that classify actions this way can automate far more, far sooner, without accepting more risk.
Frequently asked questions
Should agents ever take irreversible actions?
Rarely, and only behind strict approval gates. Most security goals can be met with reversible containment while people decide on permanent steps.
Who should be able to reverse an agent's action?
Any on-call analyst, and ideally the agent itself on request. Reversal should not depend on finding the one person who knows how.
Anomali and Reversibility of Agentic Actions
Anomali designs agent actions to be explainable and reversible, with people overseeing the decisions that matter.
Request a demo to see reversible response in Anomali.
Read next