MGnarf is an open-source execution security gate preventing AI agents from self-approving protected actions, paired with MGnarfTest, an adversarial testing project; the author is recruiting developers and security researchers to independently challenge and break the system.
"MGnarf stops AI from giving itself the green light. MGnarfTest makes sure the red light actually works." AI agents are becoming incredibly powerful. They can write code, modify systems, deploy software and execute real-world actions. But there's a fundamental problem: What happens when AI is allowed to check and approve its own work? That's the problem MGnarf is trying to solve. Two projects. One mission. 🛡️ MGnarf — The Gate An open-source execution security gate designed to prevent AI agents from executing protected actions without valid, independently verified authorization. No valid evidence? BLOCKED. Valid evidence and authorization? PASS. The agent must not be able to create its own approval. 🔥 MGnarfTest — The Challenger A separate open-source testing project designed to challenge MGnarf. Its purpose is simple: Try to break the gate. Find vulnerabilities. Write adversarial tests. Challenge assumptions. Test race conditions, identity separation and authorization bypasses. The testing system itself must also be open to scrutiny. Who am I looking for? I'm looking for developers, security researchers and curious hackers who enjoy solving difficult problems. Especially people interested in: Python development and automated testing AI agent security and execution control Adversarial testing and security research Independent verification and reproducible evidence Open-source collaboration You don't have to believe MGnarf works. In fact, I'd prefer you tried to prove it doesn't. Why join? This isn't another AI chatbot or wrapper. It's an attempt to build something underneath AI systems: a security checkpoint that doesn't simply trust the agent's own claims. MGnarf already has a working reference implementation and has undergone adversarial testing. But internal tests aren't enough. Independent developers are needed to challenge the design, expose weaknesses and help make the testing process credible. Everything is being developed openly. Contributions, test cases, failures and fixes should be visible and reproducible. Explore the repositories, open an issue, propose a test or submit a pull request. If you're interested in taking a more active development role, get in touch. The goal isn't to prove that MGnarf is perfect. The goal is to build a system that can survive people trying to prove it wrong. Because if AI is going to act independently, somebody needs to check the checker.
Arc Gate is a security tool for AI agents that tracks entire conversations to detect adversarial behavioral drift across multiple turns, unlike traditional per-message checks. The author seeks teams with real agent workflows to test it.
Arc Gate is an OpenAI-compatible firewall that tracks authority across entire AI agent sessions, escalating from allow to block before tool calls execute. It is available as a live demo and open-source on GitHub.
OpenAI outlines comprehensive security measures on the path to AGI, including AI-powered cyber defense, continuous adversarial red teaming with SpecterOps, and security frameworks for emerging AI agents like Operator. The company emphasizes proactive threat detection, industry collaboration, and security integration into infrastructure and models.