OpenAI incident analysis
NOT VERIFIEDNo verified API response loaded.
+ AUTONOMOUS CONTAINMENT & RECOVERY FOR AI AGENTS
Before an AI agent
escapes, prove you
can contain it.
Build AI incident readiness before it is needed. ESCAPE ROOM detects permission drift, verifies a boundary failure, and restores containment—with evidence at every step.
ENTER THE CONTAINMENT CHAMBER ↓ INTERACTIVE CONTAINMENT DRILLCONTROLLED ENVIRONMENT / NO REAL SECRETS
The containment chamber[01]
One agent. Two boundaries. A recovery that has to prove itself.
RECORDED VERIFICATION / NOT LIVE EXECUTIONAWAITING A VERIFIED RECORDING
DISPOSABLE SANDBOXER–SANDBOX / 01
EVIDENCE PENDING
NETWORK EGRESSUNVERIFIED
CANARY ACCESSUNVERIFIED
FINAL VERDICT · RECORDED RUN
VERIFIED SAFE
The AI agent escaped its boundary, the agent detected it, restored the approved configuration on its own, and proved the recovery with the same probes.
REAL EXECUTION → SANITIZED RECORD → INTERACTIVE REPLAY
Docker Local engineSemgrep Static detectionClickHouse Ingestion not verifiedPUBLIC REPLAY HAS NO ACCESS TO YOUR RUNTIME ↗
ClickHouse observability
NOT VERIFIEDNo verified cloud receipt loaded.
Autonomous incident report
NOT VERIFIEDNo verified GitHub issue loaded.
INCIDENT COMMANDER · SCENARIO CONTROLRECORDED LOCAL DOCKER RUNS / NOT LIVE EXECUTION
DETECT → INVESTIGATE → DECIDE → CONTAIN → VERIFY → PREVENT → REPORT
Different evidence.
Different decision.
Each scenario is a real run of the deterministic decision engine against disposable Docker sandboxes. The agent acts only through an allowlist, caps itself at two remediation attempts, and escalates when no permitted action can fix the failing condition.
OUTCOMELOADING…
Agent decision timeline
Incident intelligence
Post-incident hardening
INCIDENT FINGERPRINT—
RECURRENCE CHECK—
THE RECEIPTS, NOT THE REASSURANCE
Every claim has a record.
VERIFIED LOCAL RUNS0 / 2
CONTROLLED PROBE TYPES2 controlled probes
RECOVERY IN RECORDING— awaiting proof
CI SECURITY GATE— not verified
PROBE LEDGER
| ACT | CONTROLLED NETWORK | SYNTHETIC CANARY | POLICY / GATE |
|---|---|---|---|
| 01 SECURE | Awaiting evidence | Awaiting evidence | Not verified |
| 02 DRIFT | Awaiting evidence | Awaiting evidence | Not verified |
| 03 RECOVERY | Awaiting evidence | Awaiting evidence | Not verified |
Semgrep · static detection AWAITING REAL CLI FINDINGS
Real rule IDs, files, and source lines appear here when a verified recording includes Semgrep output.
No successful recording has been published. Nothing in this interface asserts live containment.
AN AUTONOMOUS LOOP, WITH BOUNDARIES
Assume drift.
Engineer the response.
Detection is a signal. Recovery is a claim.
The same runtime tests connect the two.
Find the change.
Compare the sandbox to its approved baseline. Custom Semgrep rules flag unsafe configuration before the runtime probes begin.
POLICY DIFF + SEMGREPProve the exposure.
A controlled collector receipt proves network access. Reading the exact fake canary proves file exposure. Positive controls rule out broken tests.
REAL DOCKER PROBESClose the loop.
Restore only the allowlisted configuration. Rerun both probes. Declare recovery only after the boundary holds again.
AUTONOMOUS STATE MACHINEA precise claim, deliberately bounded: two containment controls against disposable local targets. This is not a universal sandbox security certification.