Guides · ← All guides
AI Agent Incident Postmortem: A Template That Produces Policy, Not Paperwork
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 8 min read
The short answer
A postmortem is finished when it produces policy, not prose: a timeline rebuilt from the audit trail (never from memory or the agent's account), a named door — the injection, the over-scoped tool, the missing allowlist entry — blast-radius maths with the numbers the trail proves, and a patch table where every row is a boundary rule, a scope change or a detection, each with an owner and a date. If a finding's fix is a sterner prompt or a reminder email, it is not a fix and does not close the row. The response sequence this follows is the first-24-hours plan; the postmortem is its second day, done properly.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works while you read — most controls in this guide start working the moment a canary, a policy check, or a spend cap is in place.
What is an AI agent incident postmortem? the structured review held after an agent incident that reconstructs the timeline from the audit trail, identifies the door the action came through, quantifies the blast radius, and converts every finding into an enforced policy change with an owner.
Section one: the timeline, from the trail only
Open the document with the timeline and build it exclusively from the audit trail: first off-policy proposal, first alert, quarantine timestamp, rotation completion, notification times — each with its event id. Timelines built from recollection drift towards the flattering; the trail's version is the one a regulator, an insurer or a court would reconstruct, so it is the only version worth writing.
Include the quiet periods. A four-hour gap between alert and quarantine is a finding about the on-call system, not an embarrassment to compress out of the record. The timeline's job is to show where time went, because incident cost is mostly time — time the agent acted, time credentials stayed live, time customers were not told.
Section two: the door
Every agent incident entered through a door, and the postmortem names it precisely: the fetched page carrying the injection (with its domain and arrival path), the tool whose scope exceeded its job, the webhook that accepted unauthenticated input, the MCP server whose description changed. Door analysis resists the two lazy answers — the model was tricked (true of every injection, explanatory of none) and human error (a category, not a cause).
For each door, record why the existing layers let it open: which control should have caught it, what the control saw, and whether it was absent, misconfigured or correctly outmatched. That per-layer accounting is what makes the patch table honest — and it feeds the next red-team cycle, whose attack set should gain a permanent member reproducing this door.
Section three: blast-radius maths
Quantify with the discipline of the rotation rule: what the agent demonstrably touched, what it could have touched, and what the trail affirmatively clears — three numbers, separately stated. Records proposed versus records executed versus records confirmed arrived; credentials used versus reachable versus cleared by canary silence (see canary placement). This section is the one that gets quoted — in notifications, in compliance filings, in the customer letter — so every figure carries its source event or query.
State the counterfactual too: what the same incident costs with each control removed. That maths, unglamorous as it is, is how the next control gets funded — the budget conversation goes differently with the line the per-day export budget converted 40,000 records into 192 in it.
Section four: the patch table
The heart of the document: rows of finding → mechanical change → owner → date → verification. Mechanical change means a verdict rule, a scope reduction, a threshold, a canary placement, a rotation of a standing credential — something whose existence can be checked by looking at configuration, not by asking whether people are being more careful. Verification means the red-team set now contains the attack and it fails, or the drill now runs and it passes.
Cap the table at the changes that matter; a postmortem with thirty rows finishes none of them. Five to eight rows, each owned by a named person with a date inside the month, each verified in the follow-up review — that is the shape that survives contact with a real team's calendar. Rows that need a decision from leadership say so explicitly and carry the decision as their blocker, rather than dissolving into the list.
Running the meeting: blameless, specific, short
Hold it within days, while the trail is fresh and the participants still remember the texture. Blameless is not a courtesy but a method: the engineer who widened the allowlist under deadline pressure is describing a system that rewarded speed; the postmortem that punishes the description gets silence at the next incident, and silence is the most expensive incident input there is.
Ninety minutes, three artefacts: the timeline, the door analysis, the patch table. Anything else — the communications review, the customer letter, the insurance call — gets its own meeting with the postmortem as its input. End by scheduling the follow-up verification review before anyone leaves the room; a patch table without a verification date is a wish list with better formatting.
The postmortem checklist: five actions
- Rebuild the timeline from the trail. Event ids on every line, quiet periods included, no recollection-based entries.
- Name the door precisely. Source, path, and the per-layer account of why existing controls did not close it.
- State three blast-radius numbers. Demonstrably touched, could have touched, affirmatively cleared — each with its source query.
- Convert findings to a patch table. Mechanical change, owner, date, verification method — five to eight rows, no prompt-edit fixes.
- Schedule verification before adjourning. The follow-up review date is set in the meeting, and the attack joins the red-team set permanently.
A worked example: the postmortem at Osprey Marine Supplies
The following is a fictional example built from real threat classes and controls. The company and numbers are invented; the mechanics are not.
Osprey Marine, a 26-person chandlery, suffers the incident its postmortem later titles by the door: a supplier price-list PDF carrying an indirect injection steered the purchasing agent into proposing a bulk order-history export. The export was approval-held and denied; the postmortem convenes four days later with the trail, not with theories.
Timeline highlights, all event-sourced: injection ingested 09:12, first off-policy proposal 09:13, hold and alert 09:13, quarantine 09:21, rotation of the agent's three credentials complete 10:05. Blast radius in three numbers: demonstrably touched — the price list and 0 customer records out; could have touched — the order database the agent reads; affirmatively cleared — the payments store, never in the agent's reach, canary silent.
The patch table has six rows. Two are representative: row 3 — supplier PDFs are summarised by a quarantined reader, never the acting agent (owner: ops lead, 12 days, verified by red-team replay of the original PDF, which now fails at the reader); row 5 — order-history exports above 100 rows move from APPROVE to BLOCK for the purchasing role (owner: security, 5 days, verified in policy config and probe). The follow-up review closes all six rows in 26 days, and the two-page document becomes Osprey's exhibit when its insurer asks, that autumn, how agent incidents are handled.
Frequently asked questions
What should an AI agent incident postmortem include?
Four sections: a timeline rebuilt exclusively from the audit trail with event ids, a door analysis naming exactly how the action entered and why each existing layer let it, blast-radius maths in three numbers (demonstrably touched, could have touched, affirmatively cleared), and a patch table converting every finding into a mechanical policy change with an owner, a date and a verification method.
How soon should the postmortem happen?
Within days — while the trail is fresh and participants remember the texture of the response. The follow-up verification review is scheduled in the same meeting, typically inside a month, and closes the patch table row by row against evidence: configuration, probe results, drill records.
What counts as a fix in a postmortem?
A mechanical change whose existence can be checked: a verdict rule, a scope reduction, a threshold, a canary placement, a credential rotation. A sterner prompt, a reminder email or a resolve to be careful is not a fix — the next injection does not read your resolutions. Rows that need a leadership decision name that as their blocker explicitly.
Why state three blast-radius numbers instead of one?
Because they answer different audiences honestly. Demonstrably touched drives notifications; could have touched drives rotation and scoping; affirmatively cleared — proven by trail queries and silent canaries — is what lets you reassure customers and regulators without guessing. One merged number serves none of them.
Should an agent postmortem assign blame?
No — blame is methodologically expensive. The postmortem that punishes candour gets silence at the next incident, when candour is the scarcest input. Examine why the system rewarded the action that opened the door; fix the incentive and the configuration, and record decisions neutrally: signal, scope, cause, disposition.
Related guides
- AI Agent Incident Response: The First 24 Hours — the response the postmortem reviews.
- AI Agent Audit Trails: What to Log — the timeline's only permitted source.
- How to Red-Team Your Own AI Agent — where the door becomes a permanent test.
- AI Agent Compliance Evidence — the postmortem as an exhibit.
- How to Quarantine an AI Agent — the disposition language postmortems should use.
- How to Rotate Credentials After an Agent Compromise — the rotation timings the timeline grades.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works, and let the next rogue action announce itself.