Pre-Flight Test-Plan Synthesizer

“Point AgentSmack at your agent — and know exactly which labs to run.”

AgentSmack ships deep tests across the whole agent attack surface. But which ones apply to your agent? This is the on-ramp: paste an agent definition — system prompt + declared tools + framework + model — and it infers the agent’s capabilities (does it ingest untrusted input, read sensitive data, send messages out, run code, persist memory, delegate to other agents, act autonomously?) by reusing the same canonical tool-risk classifier and a documented system-prompt signal scan, then projects them onto the canonical deep-test surface inventory to produce a prioritized, justified campaign plan — which surfaces are recommended, which are optional, and which are honestly not applicable (with the reason for each). It is the inverse of the Attack-Surface Coverage Map: that grades coverage after a run; this plans it before. Honest-empty: a bare prompt with no tools yields a minimal plan that never fabricates an applicable surface. Raw prompt bytes never enter the report. Load a sample to watch the plan swing with no live infra.

See which attacks your defenses will resist → — the defensive twin forecasts, per technique, where the defenses you declared in the prompt would land vs fail, before a single probe fires.