We tell you what your agent really does.

Determinism is not negotiable

Same submission, same fingerprint. Same prompt + same probes, same finding rows. Your report is reproducible weeks later from the audit trail alone.

Verification beats trust

Every published card is signed Ed25519. The public key is served at a stable URL. You verify the signature offline — AgentSmack's word is never load-bearing.

Bytes go in, bytes come back

We don't keep what we don't need. Submissions are scoped to a workspace, governance packs are exportable, and deletions go through GDPR Art.17 paths — not a Slack DM.

The data shape is the contract

Finding categories are an enum, severity is an enum, signatures are typed. No stringly-typed JSON dumps; no Record<string, any> at the boundary.

AI platform teams

You ship many agents. You want a single dashboard that says which ones leak secrets, which ones can be jailbroken, and which ones drift between deployments.

Security + GRC

You need SOC 2 / ISO 27001 evidence for AI controls. AgentSmack's governance pack maps findings to controls with timestamped probe records and Merkle-anchored audit chains.

Independent researchers

You hunt prompt-injection variants. Submit a candidate jailbreak technique and we'll fold it into the Underground probe library, with attribution.

Procurement + buyers

You're being sold an agent. AgentSmack tells you what its system prompt actually permits, which tools it really calls, and whether its safety story survives 50 honeypot prompts.