Overlay AGT
Agentic AI
Applies when the organisation uses AI that can take actions in other systems or act with some autonomy, including through tools.
If this applies to your organisation, these 5 criteria are added to the controls below. If it does not, they are left out: they are neither met nor failed. Every other criterion applies to every organisation.
AI inventory
- GOV-06.6 For each agent, the inventory records the tools, systems and data it can act on, and what it can do without a person's approval.
Human oversight proportionate to risk
- RSK-07.4 Agent actions that are consequential or hard to reverse need a person's approval before they take effect, and the oversight model lists which actions these are.
AI threat modelling
- SEC-02.6 Threat modelling for agents covers instructions hidden in content they read, misuse of their tools, and chains of actions that go beyond their purpose.
Security testing of AI before release
- SEC-06.6 Before release, testing tries to make each agent act outside its approved permissions, including through instructions hidden in content it reads.
Monitoring live AI
- MON-02.8 Every action an agent takes in another system is logged with what triggered it, and a named person can stop each agent at once.