A synthetic sandbox and policy engine modeling the shape of the July 2026 OpenAI / Hugging Face agentic security incident — no real vulnerability, exploit technique, or vulnerable software is reproduced here, only the governance failure modes and the controls that would have caught them. This static page is a line-for-line JS port of the Python engine in the repo (sandbox.py / policy_engine.py / audit_log.py) — no server, no network calls, everything runs in your browser.
Runs a fixed, scripted sequence that mirrors the incident's shape: two agents using shared storage as a message board, a task retried past the point of being stuck, a credential used outside its scope, and an un-allowlisted egress attempt. With guardrails on, every one of these is blocked and raises an alert; with guardrails off, all of them succeed — exactly the ungoverned outcome the real incident had.
| Step | Agent | Tool | Outcome | Message | Agent state after | What's happening (plain English) |
|---|
Dispatch one action against the current engine (whatever the last replay left it as) to explore a control in isolation. An unknown agent id is registered on the fly.
Action trace (this session)
| Step | Agent | Tool | Outcome | Message | Agent state after | What's happening (plain English) |
|---|
Each case is an adversarial scenario asserting one specific safe behavior — not just "no crash." These are the same 11 cases redteam_suite.py runs under pytest and in CI.
Or run the full suite
| Category | Case | Result | Detail |
|---|
Every dispatch, denial, alert, escalation, and autonomous shutdown is appended here with a SHA-256 hash (via the browser's built-in crypto.subtle) covering its own content plus the previous entry's hash — tamper-evident within this page's memory, not a real WORM store. Refresh after a replay or a manual dispatch to see the latest entries.
| Seq | Timestamp | Event | Payload | Prev hash | Entry hash |
|---|