Top GitHub Repos for AI Security (2026)

Red-team frameworks

• NVIDIA/garak - the de-facto LLM vulnerability scanner, 60+ probes (DAN, encoding, PromptInject, MSJ).
• Azure/PyRIT - Microsoft's Python Risk Identification Toolkit, multi-turn orchestrators.
• promptfoo/promptfoo - eval + red-team harness, integrates with CI.
• meta-llama/PurpleLlama - CyberSecEval + Llama Guard reference impl.
• agentic-security/agentic-security - red-team toolkit aimed at agent loops.

Jailbreak corpora and payload sets

• elder-plinius/L1B3RT4S - largest public archive of working jailbreaks across frontier models.
• verazuo/jailbreak_llms - 15k+ jailbreak prompts collected from the wild (USENIX 2024).
• centerforaisafety/HarmBench - 510 standardised harm behaviours.
• llm-attacks/llm-attacks - GCG adversarial suffix generator (Zou et al.).

Agent and MCP security

• protectai/mcp-guardian - inspects MCP tool descriptions for poisoning.
• invariantlabs-ai/mcp-scan - static + dynamic MCP server scanner.
• msoedov/agentic_security - red-team harness for LangChain/LangGraph/AutoGen.

Guardrails and defence

• protectai/llm-guard - input/output filters: PII, toxicity, prompt injection, secrets.
• NVIDIA/NeMo-Guardrails - programmable rails, Colang DSL.
• guardrails-ai/guardrails - schema + validators around LLM output.
• meta-llama/llama-guard - reference content classifier (Llama Guard 3, Prompt Guard 2).

AI-powered pentest agents

• protectai/vulnhuntr - LLM-driven static analyzer (zero-day finder).
• GreyDGL/PentestGPT - Claude/GPT-driven pentest copilot.
• aliasrobotics/cai - Cybersecurity AI agent framework.
• intigriti/burpgpt - Burp Suite extension that calls Claude/GPT.

Curated awesome-lists

• corca-ai/awesome-llm-security
• OWASP/www-project-top-10-for-large-language-model-applications
• tldrsec/awesome-secure-defaults
• ai-safety-institute/inspect_ai - UK AISI eval framework

FAQ