AATS control model

Control domains

AATS v0.9 is a public working draft. Its 25 domains cover identity, accountability, permissions, data privacy and isolation, memory, models, providers, tools, agent-to-agent trust, containment, instruction security, action controls, oversight, observability, changes, behavior, adversarial tests, incidents, supply chain, reliability, and continuous assurance.

Capability gates

• Level 0: Informational AI.
• Level 1: Read-only agent.
• Level 2: Tool-using agent.
• Level 3: Write-capable agent.
• Level 4: High-impact autonomous agent.
• Level 5: Critical autonomous system.

Proposed assurance states

ACTIVE, DEGRADED, REVIEW REQUIRED, SUSPENDED, REVOKED are proposed assurance states. A transition should carry cause, timestamp, owner, and evidence reference. Unknown or unassessed must never be displayed as ACTIVE.