How is AI Control secured?
AI Control uses defense in depth: secure boot, hardware-backed identity, key rotation, credential revocation, least privilege, capability-based access, network segmentation, mutual authentication, signed software and policies, tamper-evident logs, sandboxing, continuous monitoring and independent safety controls — mapped to a documented threat model covering credential theft, prompt injection, privilege escalation, unknown AI channels, compromised robots, outages, insider misuse, misconfiguration and telemetry manipulation.
- 12 documented threat categories.
- 14 security controls.
- No single layer can grant authority on its own.
Frequently asked questions
How does AI Control protect against compromised developers?
Signed software and policies, two-person policy approval and tamper-evident audit make unauthorized changes detectable and reviewable.
How should vulnerabilities be reported?
Through the Security inquiry on the contact page.