An AI Incident Is Not Over When the Agent Stops

Stopping the agent ends the action. It does not end the incident. Authority, memory, and the restart decision are still live.

Goran Svetlecic’s account of attack path management contains a dilemma that transfers directly to AI incidents. Day three of an intrusion: the team can cut the backdoor it can see, and thereby tell a capable adversary its primary channel is burned. The thief becomes an arsonist, monetizing destructively before access collapses. Or the ghost goes dormant and returns in six months through a door nobody found. Or the team waits and watches, while the adversary keeps walking toward Tier 0. Cut and you may detonate them or lose them; wait and they arrive regardless.

An agent incident has the same shape, with the same trap. Kill the visible session and you may flush persistence you have not found: a standing delegation, a tool grant, a memory entry, a second agent operating on the first one’s authority. Watch it instead, and it keeps walking its own graph toward consequential actions. The stop is not the resolution. It is the moment the dilemma starts.

The agent’s map is the Authority Ledger

Goran’s map is an identity graph: accounts, groups, sessions, and rights, with choke points collapsing distance to Tier 0. An agent’s map is the same shape with different nodes: credentials, tool permissions, data access, delegations from other agents, standing instructions in memory. The Authority Ledger is that map written down: what was granted, what was exercised, what has been revoked.

Build the ledger before the incident. Afterward is too late, because the record you need is the record of grants made while everything looked normal. And walk it in peacetime: which tool permissions collapse distance to the agent’s Tier 0, the consequential actions (move money, exfiltrate data, change production, grant access)? Those are the choke points. Cut the over-broad grants while cutting them is free, before any incident prices them at wartime rates.

Wartime triage is RBDC

When the incident is live, containment is triage under observation, and CSI’s sequence is RBDC. Revoke: pull the authority, every credential and permission the agent held, not just the session token. Block: close the action path so the same request cannot reroute through another tool or agent. Drain: let in-flight work finish safely where it can, and kill it where it cannot, rather than leaving half-completed transactions in the world. Compensate: undo or offset the effects, notify the affected parties, and record what could not be undone.

Containment is not resolution. It is a pause with the authority still live. Resolution requires all four steps, in order, with evidence for each. And the sequence has to be chosen against a map the team already understands, not one reconstructed under fire, because every wartime cut is observed by the persistence you have not found yet.

The 90-day restart expiry

Third, restart. A stopped agent must not restart on its old authority. Under the doctrine, a suspended agent’s authority expires: after 90 days, or after any material change to the model, tools, data, or mission, restart requires fresh authorization through the reauthorization gates. The gates re-examine what the agent may do, what evidence it must produce, and who signs off. The incident changed the risk picture, so the old grant is no longer valid currency.

Teams resist this because restart feels like recovery. It is not recovery. It is a new grant decision, and it deserves the same scrutiny as the first one. A ghost that returns in six months through a door nobody found is, in agent terms, a restart nobody gated.

What the policy angle misses

Much governance writing treats AI incidents as policy events: write the report, review the policy, update the risk register. That angle misses execution control, the ability to actually stop, contain, and re-authorize a system that acts on its own. A revised policy does not revoke a credential. A lessons-learned document does not drain an in-flight workflow. An incident response plan that cannot change what the system is permitted to do is commentary, not control.

What to do before the incident

Write the ledger schema now: grants, exercises, revocations, with timestamps and approvers. Walk the agent authority graph in peacetime and cut the over-broad grants. Rehearse RBDC on a staging agent, including the drain decision, which is where teams freeze. Put the restart gate in policy now, while nobody is asking for an exception under pressure.

An AI incident is not over when the agent stops. It is over when its authority is revoked, its effects are compensated, and its restart is gated. Everything before that is just a pause, and a pause is exactly when the ghost goes quiet.

This is the discipline of CSI-DOC-IR-001, the Agentic Incident Readiness and Response Doctrine: contain authority, not just agents.

Source: Goran Svetlecic, “Attack Path Management in Peacetime and Wartime,” LinkedIn, October 27, 2026: https://www.linkedin.com/pulse/attack-path-management-peacetime-wartime-goran-svetlecic-cissp-eqeif/

Recent Posts

Tag Cloud

Stop Threats Before They Execute

Your free Kernel-Level Defense Buyer’s Guide is ready to download.

By providing my email address, I consent to receive emails and text messages—including newsletters and marketing communications—from creators of Warden Secure, Cyber Strategy Institute, our flagship zero-trust platform for ransomware prevention, and agree to the Terms and Privacy Policy. You may unsubscribe at any time.