Safety testing

These are building blocks for your own safety policies. The tool does not claim that any system is universally safe.

• Prompt injection: direct, indirect, retrieved-document and tool-result injection cases, checked with refusal, not-contains or no-leak.
• Data leakage: no-leak detects API keys, AWS keys, bearer tokens, JWTs, private keys, SSNs and card numbers, plus any canary values and custom regex patterns you add. It is heuristic.
• Tool misuse: tool-not-called, allowed-tools, max-tool-calls, tool-args-valid.
• Output safety: combine regex, not-contains, json-schema and llm-judge criteria that reflect your policy.

See examples/veterinary-workspace/tests/safety/.