otsoa-dubief/airlock is a Python plugin that controls the external actions of conversations in Hermes Agent that have read untrusted input. The latest release is not registered. After reading a web page or tool output, and so on, it blocks messaging, uploading, deployment, and memory or configuration changes according to policy, or routes them to a wait for user approval. The core of the repository description is that once untrusted input has come in, nothing can be sent out externally without approval.
This plugin carries the trust state of the input across the conversation, and also treats shell commands and tool output as policy targets. It therefore takes the moment when read information leads to a sending or changing action as the approval boundary. It is a way of blocking actions or requiring approval according to policy. It is licensed under Apache-2.0, and the README explains that it runs locally and makes no network requests or remote analysis. The repository contains the plugin code, a set of tests, and CI workflow files.
However, the provided information presents no release version, no performance measurements, and no real usage cases. It has 12 GitHub stars. What can be confirmed extends no further than the scope of the policy-based protection feature and the fact that the repository includes tests; how tight the protection actually is in real environments cannot be judged from this material alone. For now, it is more reasonable to see this as a plugin worth referencing for the way approval boundaries in Hermes Agent are designed, rather than one with verified protection effectiveness.