What is your process if an agent causes damage or exposure? Has it happened?
Clients also ask it like this:
- Do you have an incident response plan for AI agents?
- Has an AI tool ever deleted or leaked anything of yours or of a client's?
- How would you tell us if an agent exposed our data?
What the client is checking
The client wants a process that exists before the incident, and a straight answer about the past. A claim of "never" with no record behind it is a risk for you.
Three draft answers
Pick the level that is true for you today. Every statement marked [VERIFY]
must be true for your organisation before you send it. Delete a statement that is not true. Fill each
{{placeholder}}.
Not yet in place
We have no incident process written for agent incidents. [VERIFY] Our general incident process covers them as far as it applies. [VERIFY] We will add agent-specific steps by {{date}}. [VERIFY] Past incidents: {{true answer, with the period}}. [VERIFY]
Policy only
We have a written incident process that names agent incidents: {{steps}}. [VERIFY] The owner is {{role}}. [VERIFY] If your data is affected, we tell you within {{time}}, as {{contract reference}} says. [VERIFY] We have not tested the process for an agent incident. [VERIFY] Past incidents: {{true answer, with the period}}. [VERIFY]
Enforced, with evidence
We have a written process for agent incidents. We tested it on {{date}} with {{exercise}}. [VERIFY] We detect agent incidents by {{method}}. [VERIFY] The activity record shows the agent actions that pass through {{tools covered}}. [VERIFY] Gaps: {{gaps}}. [VERIFY] We would notify you within {{time}}. [VERIFY] Past incidents in {{period}}: {{number and short description}}. [VERIFY] We keep a log of incidents and near misses. [VERIFY]
Evidence to attach for the strongest answer
- The incident response plan, with its agent steps.
- The report of the last exercise.
- The incident log, with near misses.
- The notification procedure and the contract clause on notice times.
- Examples of detection: alerts or reviews of the activity record.
How to move up one level
From not yet in place to policy only
Add agent incidents to your incident plan: what counts as one, who leads, who tells the client, in how many hours. Start an incident log that includes near misses. Answer the question about the past from that log.
From policy only to enforced, with evidence
Run an exercise with a made-up agent incident. Set up detection on the activity record. Fix every gap the exercise shows and record the fix.
Where KYA fits
KYA does not give you an incident process. It gives evidence for one. The activity record shows what an agent ran and what was blocked. kya investigate runs fixed detectors over the local record, groups the hits into incidents and writes fix briefs for coding agents. Security alerts are off by default. When on, they tell you that a detector fired. They are advisory and never block a call.
A hosted workspace adds incidents across all reporting machines with a workflow (acknowledge, assign, resolve), routing rules and alert channels. Email is on all plans. Slack and Teams incident cards need Pro or higher. PagerDuty and Opsgenie need Scale or higher.
KYA sees only machines where it runs. A clean record is not proof that nothing happened. The answer to "has it happened" is a fact about your organisation. A KYA report cannot give it.
Get all 13 answers in one document
Choose the level that is true for each group. You get a preview at once and the full pack by email.
These are drafts. Send only what is true for your organisation. This is not legal advice.