Skip to content

Garak


Summary

A vulnerability scanner similar to the Metasploit framework designed to discover weaknesses or unwanted behaviors in LLMs, such as prompt injection, data leakage, hallucinations, and jailbreaking. Modular architecture allows for expansion and customization with plugins. Includes an auto red-team module.


Key Takeaways

  • LLM Red Teaming.
  • Investigation of the failure modes of LLMs and the conditions that lead to them.
  • Model robustness testing.
  • Automated reporting.
  • Write custom plugins to tailor garak to specific use cases.
  • Cross-platform support - Hugging Face, OpenAI API, Ollama, etc.

  • TBD

Additional Sources


Tags

read-team, evals, llm-security, audit-trails, prompt-injection, jailbreak, data-leakage, robustness, automation


License

Apache-2.0