HOL

An AI agent can write your next feature. It can also run one command that wipes your project, reads your cloud keys, or sends private data somewhere it shouldn't. By the time it apologizes, the damage is done. HOL Guard is open-source antivirus for AI agents. It places an independent security check between the agent's decision and the tool it wants to run. Ordinary work keeps moving. Dangerous operations are blocked. Sensitive actions require your approval, with a clear record of the decision. It runs on your machine, even offline. For this hackathon we brought that protection directly to Meta's Muse Code. We built a native Muse adapter that discovers the agent's tools, skills, plugins and MCP servers, then checks supported commands and file operations before Muse can execute them. A denied action is stopped before execution. An approved action resumes where it left off; the developer doesn't have to restart the task. Muse's model can do the work while the developer keeps control. We also raised the bar for proving safety. Guard Gauntlet gives a real coding agent ordinary tasks and synthetic attacks, then checks its actual tool calls, Guard's decisions and the effects on files or network connections. A model that refuses an unsafe instruction gets no credit as a Guard block. The tool action has to be attempted, and Guard has to stop it. Since the October 26 kickoff, the scenario catalog grew from 166 to 276 entries, and Muse gained its own installed-acceptance tests. This builds on an existing open-source product, with 2.6M+ combined PyPI downloads across HOL Guard and plugin-scanner. The Muse integration and expanded tests are public. We want developers to trust AI agents with bigger jobs, without giving those agents unlimited permission to act. HOL is an open-source organization with a growing ecosystem (see https://HOL.org/open-source), outside the context of the hackathon we'd still love and appreciate the opportunity to work with Meta to secure AI and help make the ecosystem safe.