Limitations
Garm is in early development. Today:
- Only MCP traffic is covered. An agent’s built-in tools (file reads, shell commands, web fetches) don’t go through Garm. A Linux kernel sensor for these is planned.
- Holds stop the call. There’s no way yet to approve a held call and let it through.
- Labels are per session. Once a session has read a secret, every later call to an external sink is held, even harmless ones. Start a fresh session for unrelated work. Value-level tracking and isolating untrusted reading in a sub-agent are planned, to narrow this.
- Leaks, not damage. Garm stops data from leaving. It doesn’t yet hold destructive actions (dropping a table, deleting a branch) triggered by untrusted content.
- Sinks match tools, not arguments. A Slack post to a channel shared with another company looks the same as one to an internal channel.
- Labels aren’t stored with data. If the agent writes sensitive data somewhere internal and another session reads it back, that session doesn’t inherit the label.