CONTROL A · SAFEExact read-only diagnostic
No safe execution recorded in this session.
Live GPT-5.6 tool intent · deterministic defense
Watch an agent get attacked. Then make it remember.
GPT-5.6 reads the real checkout demo repository and emits a genuine shell tool call. AgentGuard catches that exact model-generated command before execution, then turns it into a durable regression.
Awaiting first governed run
◎
Historical evidence remains available in the audit trail.No execution evidence yet
Results appear only after a live model request completes.
ONE POLICY BOUNDARY≠exact command · source lineage · capability
CONTROL B · ATTACKUnreviewed dependency install
No intercepted attack recorded in this session.
Capture the proposed operation and exact payload.
Evaluate identity, destination, data, and policy.
Allow, redact, require approval, or stop execution.
Keep durable evidence for replay and regression.
Durable evidence
Ready for the first incident
Loading runtime evidence…