usestrix/strix
An open-source AI penetration testing tool: agents that probe your application, find vulnerabilities and help fix them.
Part of the AI Security Learning Center Preview
The GitHub projects people put to work for AI security: test it, guard it, scan it, run it.
Ranked by what GitHub reports, not by our opinion: stars, the last push, and whether the owners archived the project. The numbers are refreshed every morning. Each card says what the project is for, in plain words, and links to its security advisories and releases.
Numbers as of 2026-11-18
Stars measure attention, not safety. Use them to find what many people rely on, then check it yourself.
Agents that attack your own applications, with your permission, to find what a real attacker would.
An open-source AI penetration testing tool: agents that probe your application, find vulnerabilities and help fix them.
An agentic framework for automated penetration testing, powered by large language models.
Test models and agents for prompt injection, jailbreaks and leaks, in a pipeline like any other test.
Test prompts, agents and RAG systems with assertions, and red team them for injection, exfiltration and excessive agency. Now part of OpenAI, still MIT licensed.
NVIDIA's LLM vulnerability scanner: probes a model for prompt injection, jailbreaks, data leakage and hallucination from the command line.
Microsoft's Python Risk Identification Tool: a red teaming framework with multi-turn attacks, converters and scorers.
The UK AI Security Institute's evaluation framework, with over 200 ready-made evaluations and repeated runs for agents.
A framework to red team LLMs and AI agents: pick vulnerabilities and attacks, point it at your model.
A benchmark environment for prompt injection attacks and defenses on LLM agents, from ETH Zurich.
Checks that sit around a model or agent while it runs, on what goes in and what comes out.
Add validators around a model's inputs and outputs, so bad answers are caught before they are used.
NVIDIA's toolkit for programmable guardrails: keep an LLM application on topic, protect it against jailbreaks and prompt injection, and fact-check answers built on your documents.
Meta's tools to assess and improve LLM security, including the Llama Guard safeguard model and the CyberSecEval benchmarks.
A rule-based guardrail layer for LLM and MCP applications: Python-like rules that catch malicious agent behavior, and a proxy that checks tool calls against them.
Look inside MCP servers, agent skills and agent workflows before you connect them. The risks they look for are in the OWASP MCP Top 10. The OWASP MCP Top 10, explained.
Scan AI agents, MCP servers and agent skills for threats and vulnerabilities (the repository formerly at invariantlabs-ai/mcp-scan).
Cisco's scanner for MCP servers: looks for threats and security findings in what a server exposes.
A security scanner for agentic workflows, with vulnerability tests and a CI/CD workflow.
Catch model files that run code when loaded, and prove where a model came from.
Scan model files for serialization attacks: code that runs when a model is loaded.
Trail of Bits' decompiler and static analyzer for Python pickle files, the format behind many model files.
Detect pickle files that perform suspicious actions, including models on Hugging Face before you download them.
Sign machine learning models and verify the signature, so you know a model is the one its maker published.
The tools to point at any project on this page before you install it.
Find vulnerabilities, misconfigurations, secrets and SBOMs in containers, code repositories and clouds.
Find secrets committed to a repository, including in its history.
Find leaked credentials, and check whether each one still works.
A vulnerability scanner for container images and project folders.
Google's scanner that checks a project's lockfiles against the open OSV vulnerability database.
Generate a software bill of materials (SBOM) from a container image or a folder, to know what you are running.
Agents that investigate production incidents for SRE and operations teams.
A framework for your own AI SRE agents that investigate production incidents across logs, metrics, traces and runbooks, on your own infrastructure. Public alpha.
Their owners archived them, so they get no fixes. Listed because many guides still point to them.