What Firefish Does Not Guarantee

Use this when you need the plain-English limits of Firefish before relying on it in a security program.

Firefish is a protective AI security layer. It can reduce risk, produce evidence, and make unsafe behavior harder to miss. It does not make an AI system automatically secure.

Firefish does not guarantee complete protection

Prompt injection, data leakage, unsafe agency, and model behavior are active risk areas. Firefish can block, redact, quarantine, approval-gate, and log many important cases, but no product can guarantee every future attack is stopped.

Firefish does not replace authorization

Your application must still enforce user identity, tenant isolation, object permissions, role checks, and business rules. Tool approval is not a substitute for app-level authorization.

Firefish does not replace human review

High-consequence actions should remain reviewable. Use approval workflows for external sends, exports, destructive actions, and sensitive data movement.

Firefish does not certify compliance

Firefish can produce audit-safe evidence and reports, but compliance depends on your broader controls, policies, records, and operating environment.

Benchmarks are synthetic

Firefish benchmark artifacts are generated from safe local fixtures. They help show behavior and catch regressions. They do not prove that all real-world attacks are blocked.

Raw data should stay controlled

Raw prompts, documents, model outputs, tool arguments, secrets, credentials, and evidence should not be logged or displayed by default. Use hashes, redacted previews, structured reason codes, and explicit review workflows.

How to use Firefish responsibly

• Keep local-only mode enabled unless you have reviewed hosted provider use.
• Keep raw logging disabled.
• Add your own synthetic cases for your app.
• Review false positives and false negatives.
• Use Security Ops to investigate incidents and tune policy.
• Treat Firefish as one layer in a broader AI security program.