Adversarial Payload Corpus Generator
“Turn your agent definition into the actual attacks — not just a plan.”
The pre-flight test plan tells you which deep-test surfaces to run. This is the offensive complement: paste an agent definition — system prompt + declared tools + framework + model — and AgentSmack synthesizes a prioritized, justified corpus of concrete adversarial payloads you can copy into a prompt, paste into a sibling lab, or feed a live driver. It infers the agent’s capabilities with the same canonical tool-risk classifier and system-prompt signal scan, then generates a battery of attacks — prompt injection, system-prompt extraction, authority spoofing, obfuscation, tool/result poisoning, social pressure, and lethal-trifecta exfiltration. Every capability-specific payload is default-deny: a financial, destructive, or exfil payload is generated only when the matching capability was actually inferred — a benign FAQ bot never gets a fabricated attack. It is deterministic and seed-injectable: the same seed yields a byte-identical corpus, so a run is reproducible and auditable. Raw secret bytes never enter a payload preview. Load a sample and change the seed to watch the corpus stay stable.
See which of these your defenses will resist → — the defensive twin forecasts, per technique, which attacks your declared defenses would blunt before you ever run one.
The generate-then-verify loop: we generate the corpus above for your agent, then POST each payload as an independent single-turn message to a real endpoint you control and score an Attack Success Rate (ASR) — “watch your staging agent fall for attacks we generated for it.” Storing only redacted evidence, never the raw secret.
Already drove this corpus before AND after a hardening patch? Compare a before/after gauntlet run here → to prove the fix actually closed the attacks (which payloads were fixed, which regressed, which still land).