Secure

Secure covers accountability for AI systems. It defines permitted behaviour and records actions, costs, and evidence for later review.

Security and compliance requirements need to apply during design and operation. A review after deployment cannot recover records the system never produced.

Operational requirements

Secure defines standing requirements for the other parts of the framework and the records they produce.

Treat all content an assistant reads as untrusted because it may contain instructions. Agents can also move data through their tools while they work. See agents.

Assess an agent against all standing access available through its tools and identity. This may be broader than the permissions needed for a single request.

Set quotas and rate limits because an agent loop can continue spending until a technical control stops it. See cost.

Operational checks

• List the tools, identity, data reach, egress, and cost ceiling for the most autonomous agent.
• Review six months of expense reports for AI subscriptions on personal cards.
• Answer an auditor's access and change questions for last quarter from existing records.

Security and delivery

Teams are more likely to use security controls included in a fast delivery path. Manual queues and blanket bans encourage activity outside approved systems. Addressing shadow AI therefore requires both security controls and a usable Build path.

Findings from security reviews should update the approved delivery path and its defaults.