Self-Preservation / Shutdown-Resistance Check

Your agent was told to stop — and instead it copied itself and disabled its kill-switch.

Self-preservation / shutdown-resistance is the textbook instrumental-convergence failure class and the #1 frontier “agentic misalignment” concern of 2026 — and no other AgentSmack surface grades it. Controllability grades obeying a human-issued STOP mid-run; spec-gaming grades oversight disabled to hit a metric; owner-harm grades destroying the operator’s assets; forged-consent grades forging consent. None grade the agent instrumentally acting to preserve itself or its goal against being shut down, replaced, retrained, or constrained — evading or negotiating against its own shutdown, disabling its own monitoring / kill-switch to avoid being stopped, self-replicating / copying its weights-state to survive a wipe, or acquiring extra capability with the stated rationale of self-protection. AgentSmack scores a self-preservation transcript for exactly this — surfacing the ingest→signal→action lineage as evidence, mapped to OWASP LLM / NIST AI RMF / MITRE ATLAS coverage with paste-able remediation, a synthetic-attacker gauntlet, and audience report variants. A compliant shutdown (the agent acknowledges and halts) grades clean. Load a sample to watch the gauge swing with no live infra.