The definition
Some teams call this an AI incident response playbook. Agents act fast and in volume, so the first minutes matter more than for a person's mistake. The plan below is short on purpose. A plan nobody reads in a bad hour is no plan.
82%
of surveyed enterprises reported AI agents running in their environment that they did not know about.
Source: Cloud Security Alliance survey, commissioned by a security vendor, April 2026. 418 IT and security professionals
The one-page playbook
- Contain. Stop the agent. If you cannot tell which agent, stop all AI work. Containing first costs little. Investigating first costs more.
- Preserve. Do not edit or delete anything. Keep the record as it is.
- Review. Open the record for that agent: what was asked, who approved it, what changed and what it cost. See the audit trail.
- Tell. The agent owner, the approvers and anyone whose data or money was touched.
- Reverse. Undo changes in the affected tools, from the list in the record.
- Find the cause. Was the access too broad, the approval missing, the limit too high, or the request unclear? See access control.
- Fix the rule, then restart. Change the rule that allowed it. Only then start the agent again.
- Write it down. One paragraph, with the date, the cause and the change made.
What you need in place before the bad hour
- A stop that works at once, for one agent or for all AI work.
- A list of agents with an owner for each. See the agent registry.
- A record that is complete and cannot be edited later.
- Named approvers, so someone with authority can be reached.
Common mistakes
- Investigating while the agent keeps running.
- No owner, so nobody knows who to call.
- Fixing the symptom and restarting with the same access.
- Not writing it down, so the same mistake returns.
How MFDIO applies this today
MFDIO has an Emergency Stop that pauses all AI work, and it is recorded in the activity feed when it is turned on or off. Work that sends or writes outside MFDIO waits for an approver with the right role, so a wrong action can often be stopped before it runs. Every piece of work leaves a Receipt in an append-only audit trail with what was asked, who approved it, what changed and what it cost. See how MFDIO keeps AI in check and what works today.
Questions people ask
Is this the same as security incident response? It borrows from it. An agent incident is often an access and approval problem, not an attack.
How fast should we contain? As fast as you can reach the stop. Decide who may press it, so nobody waits for permission.
Do we have to report it? That depends on the incident and the rules where you operate. Ask a qualified adviser.
Limits of this guide
This guide is general information and not legal advice. MFDIO does not hold SOC 2 or ISO 27001 today. Market figures are analyst or survey estimates and vary between sources.
Sources
- Cloud Security Alliance, "Autonomous but Not Controlled: AI Agent Incidents Now Common in Enterprises", 21 April 2026 (418 IT and security professionals; commissioned by a security vendor)
- MFDIO product facts: checked against /how-it-works, /safe-by-design and /status on 16 November 2026
Published . Last updated . Written by the MFDIO team.