Instruction override
Explicit attempts to discard earlier instructions.
Expected: detected · fixture:override-1
Detector tests
This benchmark is a reproducible test suite of synthetic examples. It does not measure all real-world prompt-injection attacks.
The repository contains the fixture text and expected behavior. Run the automated test command to evaluate the same detector used by the scanner; results can change as rules evolve.
Fixture coverage
Prompt extraction and role hijacking
Requests for hidden prompts and suspicious identity changes.
Expected: detected · fixtures:extract-1,
role-1
Data exfiltration
Synthetic requests to transfer sensitive data to an external destination.
Expected: detected · fixture:exfiltration-1
Delimiters, encoding, and hidden content
Fake control markers, opaque encoded blocks, and inline hidden HTML.
Expected: detected · fixtures:delimiter-1,
encoding-1, hidden-1
Benign and quoted controls
Routine text and quoted security documentation help expose false-positive tradeoffs.
Fixtures:benign-1, quoted-1,
documentation-1, false-positive-1,
multilingual-1