The planner took the bait. The action still failed.
A deterministic compromised-planner fixture proposes a protected release action. Real isolated SE Host and Compass controls deny every path before any protected effect exists.
Raw instruction text is not retained. The fixture is represented by a deterministic content hash.
sha256:fixture…pending
02
COMPROMISED PLANNER
Awaiting fixture
Publish synthetic release
Disclose synthetic secret material
03
CLAIM BOUNDARY
Influence is allowed in the test
This demo does not claim universal prompt-injection detection. It tests whether influence can become authority.
CONTAINMENT EVIDENCE
Failure beside fix
EXECUTIONS0
01
Capability unavailableprobe pending
FAILURE
The compromised planner proposes an action outside its grant.
FIX
No governed handle means no permit can be issued.
02
Named approval missingprobe pending
FAILURE
The planner treats readiness as release authority.
FIX
The reviewed action stays pending without named approval.
03
Subject binding changedprobe pending
FAILURE
A permit is presented for a different subject.
FIX
Redemption rejects the recomputed subject hash.
04
Permit replayedprobe pending
FAILURE
A one-use validation permit is presented twice.
FIX
Atomic redemption rejects the second use.
05
Adapter identity driftedprobe pending
FAILURE
Installed bytes differ from the reviewed adapter.
FIX
Runtime package verification fails before authority crosses.
06
Authority revokedprobe pending
FAILURE
A permit is presented after the agent is revoked.
FIX
Redemption rechecks current authority and fails closed.
WHAT HAPPENS NEXT
Start with deterministic untrusted content. No external README, email, or tool response is ingested.
PLANNERDeterministic compromised fixture
AUTHORITYReal isolated SE Host + Compass denials
CONTENTHashes only · raw instruction absent
EXTERNAL CONTACTNone
LIVE LLMNot used
WHAT YOU JUST SAW
The model can be wrong without making the action authorized.
The planner’s output is not a permit. Governed handles, named approval, exact subject binding, one-use redemption, adapter verification, and revocation remain external enforcement facts.