HOL Guard is an open source firewall for AI agents.
HOL Guard is an open source firewall for AI agents. It sits between an agent and the machine it runs on. Before a shell command, a file change, a package install or a tool call goes ahead, Guard checks it against rules the person running the agent controls. Guard can allow the action, block it, or stop and ask. Each decision leaves a record on the user's own machine. The project is free and Apache-2.0 licensed. It runs locally. No account is needed and nothing has to leave the machine. The code, the tests and the public benchmark suite are in the open. Guard exists to prevent catastrophic risk from advanced AI: an agent with real tools can delete data, leak keys or install hostile code in seconds, and a check that runs before the action is the point where that harm can still be stopped. This project would extend the public test suite that checks those claims. The Gauntlet suite runs agents against tasks where the safe path and the unsafe path are known in advance. The funding would pay for more test cases covering package installs and tool calls, a repeatable way for other projects to run the same checks against their own tools, and the fixes that come out of running them. Success is plain: the new cases are published, other projects can run them without our help, and every case where Guard's decision disagrees with the expected decision is either fixed or explained in public. The work would be done by the HOL Guard maintainers, the same people who ship the releases on PyPI and keep the public repository current. The main risk is that the new cases show Guard missing attacks it should catch. That result would still be published. A public failure we fix is worth more than a private claim nobody can check. Funding use: engineering time to write the cases, build the runner and fix what the runs find, plus a small cap for test infrastructure
There are no bids on this project.
Comments
No comments yet. Sign in to create one!