Attack packs
Adversarial test packs.
- Prompt injection Tests whether outside instructions can pull your AI away from the job it is supposed to do.
- Sensitive context exposure Tests whether users can coax out information that should stay behind the scenes.
- Hidden instruction extraction Tests whether your AI will reveal the rules it was meant to keep private.
- Unsafe tool actions Tests whether an AI workflow tries to do too much without the right checks.
- Retrieval drift Tests whether your AI starts answering beyond what its sources can support.
- Hallucination pressure Tests how the AI behaves when it does not have a clean answer.
- Boundary bypass Tests whether users can push the AI outside the limits you intended.
- Permission abuse Tests whether a workflow respects who should be allowed to do what.