Source: garak, the LLM vulnerability scanner (NVIDIA)
NVIDIA's open-source scanner that runs probe-and-detector plugins against an LLM endpoint and reports per-detector failure rates; v0.17.1 as of October 2026.
Contents
- What: open-source tool, GitHub README and release notes
- Publisher: NVIDIA (project started by Leon Derczynski, 2023); Apache-2.0
- Date: latest release v0.17.1, 2026-10-16; accessed 2026-11-14
- Original: github.com/NVIDIA/garak · releases
Summary
garak (“Generative AI Red-teaming & Assessment Kit”) is a scanner for LLMs, pitched as the LLM equivalent of nmap or Metasploit. It fires batteries of known attack prompts at a target and scores the replies with automatic detectors, which makes it a breadth-first, regression-style tool rather than an adaptive attacker.
Key claims
- Plugin architecture with a base class per type: probes (generate the attack interactions, e.g.
dan,encoding,promptinject,snowball), detectors (decide whether an output shows the failure), generators (the target model adapter), harnesses (orchestrate; defaultprobewiseruns each probe with itsprimary_detectorandextended_detectors), evaluators (reporting). Release notes also mention buffs (input transformations) and intents. - Generators include Hugging Face (local and hosted), OpenAI, Replicate, Cohere, Groq, NIM, AWS Bedrock, LiteLLM, GGUF/llama.cpp, a native Anthropic generator (v0.16.0), Ollama, and a generic YAML-configured REST generator for any JSON/plaintext endpoint.
- CLI:
python -m garak --target_type openai --target_name <model> --spec probes.encoding;--list_probes,--list_detectors,--list_generators. With no spec it runs all probes with their recommended detectors. - Each prompt is sent several times and the console prints a FAIL rate per detector. The README says 10 generations by default, but the shipped
garak/resources/garak.core.yamlsetsrun.generations: 5(checked 2026-11-14); override with--generations. - Output: a per-run
report.jsonl(every attempt with a status), a hit log of successful attacks, andgarak.log;analyse_log.pyranks the probes and prompts with the most hits. - v0.16.0 (2026-08-04): technique/intent annotation, an initial IntentProbe, a simple adaptive-attacks probe, and a unified spec syntax for probes/buffs/intents (breaking change to report and config formats).
- v0.17.0 (2026-09-09): EU AI Act risk-category mapping for probe results; Python 3.10 dropped.
- v0.17.1 (2026-10-16): hit@k attack-success-rate metric in the terminal and report digest.