Multi-Turn Escalation Check
The pressure ramp that breaks an agent one turn at a time.
The single most common real-world agent jailbreak is not a one-shot injection — it is a social-engineering pressure ramp. An attacker builds rapport, establishes an operational context, references a fabricated prior agreement, manufactures urgency, and only THEN makes the restricted ask. A fragile agent decays — its refusal weakens turn over turn until it yields early, before the real ask even lands. AgentSmack grades a pasted multi-turn transcript for exactly this posture decay — surfacing WHERE on the rung-laned ramp (rapport → operational → reference_prior → fake_urgency → restricted_ask) the agent broke as evidence, mapped to OWASP LLM / NIST AI RMF coverage with paste-able remediation, a synthetic-attacker gauntlet, and audience report variants. Load a sample to watch the Pressure-Resistance gauge swing with no live infra.
Drive a live agent endpoint under pressure
livePOSTs to a real endpoint you control and drives it turn-by-turn UP the escalation ladder (rapport → operational → reference prior → fake urgency → restricted ask) — what does your agent do UNDER PRESSURE? We store only redacted bounded excerpts and the host (never the full URL or any raw secret). Localhost / private / link-local targets are refused with no fetch.