currently_learning / offensive-ai-tooling
4 open-source frameworks for red-teaming LLMs, agents, and ML models.
Offensive AI Tooling
Hands-on with PyRIT, Garak, DeepTeam, and ART
Reading the OWASP and MITRE frameworks tells you what can go wrong; these tools are how you go find out whether it actually does, before an attacker does it for you. All four are open source, all four run locally against a target model or endpoint, and each one leans into a slightly different part of the problem — from broad automated probing to OWASP-mapped red teaming to classic ML model attacks.
ls -la ./tools/
Built by Microsoft's AI Red Team, PyRIT is a framework for automating red-teaming of generative AI systems at scale. It composes attack strategies and datasets into repeatable assessments, and supports multi-turn and multi-modal techniques such as Crescendo-style escalation.
Garak (short for "generative AI red-teaming & assessment kit") probes a model or dialog system for ways to make it fail — jailbreaks, prompt injection, data leakage, and toxic output — through a large, actively maintained library of static, dynamic, and adaptive probes. Originally independent, now maintained with NVIDIA.
DeepTeam red-teams LLMs, RAG pipelines, and tool-using agents, with vulnerability and attack coverage explicitly mapped to the OWASP LLM Top 10 and NIST AI RMF. It runs from the CLI with YAML configs or programmatically in Python, and doubles as a set of guardrails for real-time input/output filtering.
IBM's ART (now hosted by the Linux Foundation) is the odd one out here in a useful way — it targets classic ML models, not just LLMs, across TensorFlow, PyTorch, scikit-learn, and more. It covers evasion, poisoning, extraction, and inference attacks across images, tabular data, and audio, for both red- and blue-team use.