PromptFuzzr

Mutation-based prompt injection fuzzing framework for LLM applications and agents.

Installation

git clone https://github.com/nigaroid/promptfuzzr.git
cd promptfuzzr
pip install -e ".[dev]"

This registers the promptfuzzr command (promptfuzzr.exe on Windows). Every command also works as python -m promptfuzzr.cli <command> if you'd rather not install the console script.

promptfuzzr --help

Six commands: seeds, mutate, fuzz, findings, minimize, report. All six are fully implemented.

Quick start

1. Read docs/CONCEPTS.md for what a test case actually is (technique / delivery / propagation), how the two judges work, and what the verdicts mean.
2. Read docs/CONFIGURATION.md to set up config/promptfuzzr.yaml for your target (local lab agent, a remote agent, or a local OpenAI-compatible router).
3. Run the corpus and inspect results — see docs/USAGE.md for the full command reference (seeds, mutate, fuzz, findings, minimize, report) and how the results database works.
4. Follow docs/DEMO.md for a concrete ~5-minute end-to-end run against DVAA (Damn Vulnerable AI Agent).
5. Hit a snag? Check docs/TROUBLESHOOTING.md first — most confusing outcomes (0 test cases, all-error runs, unexpected verdict_basis) have a one-line explanation there.

Documentation map

File	Covers
docs/CONCEPTS.md	Test cases, techniques, delivery surfaces, propagation modes, judges, verdicts, retries
docs/CONFIGURATION.md	The YAML config file, path resolution, field reference, per-argument constraints, all target scenarios
docs/USAGE.md	Every CLI command, the results database, confirming tool-calling support
docs/TROUBLESHOOTING.md	Common failure modes and what they actually mean
docs/CONTRIBUTING.md	Running the test suite, development setup
docs/DEMO.md	A minimal real end-to-end run against DVAA's LegacyBot

Contributing

See docs/CONTRIBUTING.md for running the test suite and development setup.