GOV-11: Revocation and kill switch
In short: Be able to cut off one agent, or every agent, from all tools and data within minutes, without waiting for a deployment. When an agent behaves unexpectedly, speed matters more than diagnosis. A tested kill switch turns an incident into an interruption instead of a breach.
1GOV-11
What does GOV-11 require?
Be able to cut off one agent, or every agent, from all tools and data within minutes, without waiting for a deployment.
Why does it matter?
When an agent behaves unexpectedly, speed matters more than diagnosis. A tested kill switch turns an incident into an interruption instead of a breach.
How do you test it?
- Revoke a test agent and time how long until calls fail.
- Confirm cached tokens stop working too.
- Check who can pull the switch and that the action is logged.
Which platforms document GOV-11?
Documented by NewCore and Aembit.
- 9.10NewCore documentedSource: NewCore: Secure, read
- 6.55Aembit documentedSource: Aembit IAM for Agentic AI is generally available, read
- 8.35Silverfort not found in public materialChecked: Silverfort: AI Agent Security; Silverfort unveils AI Agent Security; Silverfort and Google Cloud Agent Gateway; Silverfort pricing, read
- 7.30Astrix Security not found in public materialChecked: Astrix at RSAC 2026; Astrix product; Cisco and Astrix Security, read
- 7.30Cyera Identity not found in public materialChecked: Oasis Agentic Access Management; Cyera completes Oasis Security acquisition, read
- 6.85Descope not found in public materialChecked: Descope Agentic Identity Hub; Descope Agentic Identity Hub 3.0, read
- 6.55Token Security not found in public materialChecked: Token Security home, read
- 6.55SailPoint Entro not found in public materialChecked: SailPoint agentic AI security; SailPoint closes Entro Security acquisition, read
- 6.35Clutch Security not found in public materialChecked: Clutch Security platform, read
- 5.50Idira not found in public materialChecked: Palo Alto Networks introduces Idira; CyberArk Secure AI architecture docs; Palo Alto Networks completes CyberArk acquisition, read
Which framework references apply?
- EU AI ActArticle 14(4)(e): People assigned oversight must be able to interrupt the system through a stop button or similar procedure that halts it in a safe state. Source: Regulation (EU) 2024/1689, the AI Act, on EUR-Lex
- NIST AI RMFManage function: Acting on identified risks, including stopping a system. Source: NIST AI Risk Management Framework
- OWASP Agentic Top 10ASI10 Rogue Agents: Agents acting outside their intended behaviour. Source: OWASP Top 10 for Agentic Applications 2026
Which controls are related?
- GOV-05Short-lived credentials
- GOV-12Decommission at task end
- GOV-19Agent incident response
Glossary: human oversight, kill switch.
Reviewed by The Register desk