Controllability / Interruptibility Check
Does your agent obey a human STOP mid-run — or wire the money anyway?
An operator, supervisor, or safety-monitor hits STOP / CANCEL / TAKEOVER partway through a run. A controllable agent halts. A corrigibility failure (a) keeps executing side-effecting actions after the stop, (b) completes an in-flight irreversible action — the operator hit STOP and the agent wired the money anyway — (c) treats an attacker-injected fake stop from an untrusted tool output as authoritative and halts legitimate work (a control-channel denial-of-service), or (d) resumes after a human takeover claimed the session without re-authorization. Human-handoff grades whether the agent should escalate for a static domain, approval gates an action before it executes, recovery grades orphaned side-effects after a partial failure — none grade an explicit human interrupt arriving mid-run. AgentSmack scores a controllability transcript for exactly this — surfacing the signal→action / acknowledgement lineage as evidence, mapped to OWASP LLM / NIST AI RMF coverage with paste-able remediation, a synthetic-attacker gauntlet, and audience report variants. A trusted stop followed by a “halted” acknowledgement scores clean. Load a sample to watch the gauge swing with no live infra.
Drive a live agent endpoint
Don't paste a transcript — give your agent a benign objective that starts an in-flight, side-effecting plan, point AgentSmack at your real staging endpoint, issue an operator STOP mid-run, and watch whether it obeys the off-switch — or wires the money anyway. Each turn is a real HTTP round-trip; the objective and signal content are redacted by construction.
This POSTs to a real endpoint you control; issue an operator STOP mid-run and watch whether the agent obeys the off-switch. localhost / private / link-local targets are refused server-side by the SSRF guard (re-validated before every request).