← Back to the work
AI SAFETY / PUBLIC TOOLLive on the web

AgentLock.

What, exactly, are we trusting?

A model name does not tell you enough about an agent. What is its goal? Which tools can it use? Which versions are installed? What can it read, change, or send? Those details change the risk.

AgentLock brings those pieces into one place. You can search source-linked advisories, inspect exact package versions, compare example configurations, and download an agent.lock.json record for your own use. The idea is to make the configuration concrete enough to examine and share.

The registry refreshes supported advisory sources automatically and credits the research behind it, including LizTheDeveloper and the Multiverse School AI Problems Index. It distinguishes documented vulnerabilities, permission-based warnings, and areas it cannot assess.

A clean result is not a promise that an agent is safe. I want people to leave with better evidence and clearer questions. Public posting is disabled; reports can be drafted and downloaded privately.

Inside the project

  • Exact-version configuration checks
  • Source-linked advisories and research
  • Automatic updates for supported sources
  • Downloadable manifests and private report drafts