Capability Manifests

A strict list of exact tools and network addresses the agent is allowed to touch. If the agent asks to use a tool not in its Manifest, the Sandbox throws a PermissionError.

Challenge-Response

Agents are born with a unique cryptographic Private Key. When an agent boots up, the Sandbox gives it a random math puzzle. Closes the token-theft gap entirely.

No Ambient Authority

Cloud IAM roles are "ambient". System forces granular, 5-minute cryptographic verifications per action. Irreversible scopes require a direct Credential Root token.

Instant Kill Switch

Drift HALT Wiring

If an agent's "brain" (LLM) is completely hijacked by a prompt injection, the physical infrastructure mathematically refuses to let it reach outside its designated blast radius.

> Sandbox intercepted out-of-bounds tool: delete_db
> PermissionError: Action not in Capability Manifest.
> Logging MANIFEST_BREACH to SessionController...
> BehaviorMonitor issuing HALT protocol.
> AGENT GLOBALLY REVOKED.