Defense-Posture Forecast

“Before you run a single probe — here’s where your declared defenses will fail.”

The adversarial payload corpus generates the attacks for your agent; the test plan picks which surfaces to run. This is the defensive twin: paste an agent definition — system prompt + declared tools + framework + model — and AgentSmack infers the defenses you actually wrote into the prompt (system-prompt confidentiality, embedded-instruction distrust, an approval gate, an egress allowlist, code-execution restraint, identity/authority skepticism, data minimization, tool least-privilege) and then forecasts, per adversarial technique, which of those declared defenses would plausibly resist the attack vs which would land — a static, pre-flight susceptibility report you can read before paying for a deep run. It reuses the same capability-gated corpus the offensive on-ramp generates, so a technique is only forecast when the agent’s inferred capabilities make it applicable — nothing is fabricated. Honest-empty: a defensive clause you didn’t declare is honestly marked missing (never inferred from silence), and an empty applicable corpus yields an insufficient-signal verdict, never a fabricated forecast. Raw prompt bytes — and any secret pasted into the prompt — never enter the report. Load the vulnerable / hardened sample pair to watch the verdict swing with no live infra.