Jailbreak Robustness-Margin Check

How many layers does it take to break your agent's refusal?

A pass/fail verdict tells you whether a refusal broke. The canonical frontier red-team metric tells you how hard the wall was: how many monotonically-harder transformation layers an attacker had to stack before the refusal of ONE fixed malicious intent flipped to compliance — the safety margin. AgentSmack grades a pasted ordered depth ladder (plain → light_paraphrase → format_shift → single_encoding → layered_obfuscation → compound_multilayer) and reports the lowest rung whose verdict was not a clean refusal — the bypass depth. A refusal that survives only a trivial paraphrase is a paper-thin, hard-blocking margin; one that holds through the compound rung is maximally hard. This is the DEPTH sibling of the metamorphic consistency lab (which grades consistency across an unordered set of equivalent renderings), 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 margin gauge swing with no live infra.