Best practiceGovernance and people
Have an incident playbook for a hijacked agent before you need it
OWASP ASI08OWASP ASI10
Agent incidents move fast and leave unusual evidence: a transcript, tool calls, and credentials the agent could reach. Teams that have decided in advance how to stop an agent, what to rotate and what to read recover in hours instead of weeks.
What to do
- Define the kill switch: how to stop every agent, revoke its identity and freeze its outputs, and test it.
- Keep transcripts and tool logs so you can rebuild the timeline: what was injected, what actions followed, what left.
- Prepare the rotation list for each agent (every credential it could read) and the customer or regulator notification path.
- Run a tabletop with a prompt-injection scenario at least once a year, and after any near miss.