nvidia/garak is a free command-line tool that checks whether LLMs and conversational systems fail in unwanted ways. The repository description defines the tool as an LLM vulnerability scanner, and the latest public release is v0.17.1. However, this record does not retain the change items for v0.17.1, so newly added features cannot be confirmed.
garak is short for Generative AI Red-teaming & Assessment Kit. It combines attack probes with detectors to test models with static, dynamic, and adaptive (static, dynamic, adaptive) probes, and the detectors judge the responses. The checks cover hallucination, data leakage, prompt injection, misinformation, harmful generation, and jailbreak, among others. It targets harmful output, information leakage, and prompt vulnerabilities together. It is written in Python and released under the Apache-2.0 license, and the repository has 9627 stars. The repository was created on May 10, 2023.
However, no benchmark comparing detection accuracy and no independent evaluation are given. The only figure confirmable in this record is a momentum of 4.4, and freshness, technology, and quality scores are not calculated. There are many checks, but no evidence of accuracy remains. Actual detection performance needs to be verified separately.