Trial
Update 07/2026: Garak has kept pace with the shift to agents and is more relevant than ever for red-teaming LLM applications:
- v0.15 (May 2026) added an Agent-Breaker probe (attacks against the tools of LLM agents), a system-prompt-extraction probe and a multi-turn GOAT probe; v0.14 brought HTML reports and JSON config (releases).
- NeMo-Guardrails support and improved refusal detection.
For radar readers building agents: garak now covers parts of the OWASP Agentic Top 10 attack surface, not only classic prompt injection.
Updated link: GitHub Repository (the leondz/garak link in the previous entry is outdated).
Trial
Garak is NVIDIA's open-source security testing toolkit for evaluating Large Language Models (LLMs). It provides a comprehensive suite of tests to assess model vulnerabilities, detect potential security issues, and ensure robustness across different scenarios.
Key Features
- Extensive test suite for LLM vulnerabilities
- Automated prompt injection detection
- Model robustness assessment
- Customizable testing scenarios
Business Value
- Early detection of security risks
- Compliance with AI security standards
- Protection against model exploitation
- Enhanced AI system trustworthiness