Hey folks, hoping to tap the collective wisdom here. I’ve been self-hosting an agent framework (not Open Claw, I know, I know — let’s call it “Project Nebula” for this thread) for a few weeks now. The setup is mostly in Docker Compose, behind a Tailscale subnet router.

Everything *seems* to be working, tasks run, results come back. But my audit logs are absolutely **flooded** with ‘denied’ entries from the agent process itself. We’re talking hundreds of lines an hour. The entries look like attempts to access certain system calls or paths that are being blocked by the seccomp/apparmor profile I have in place.

My threat model for this setup is a compromised task/plugin: I want to prevent breakout to the host or to other containers, and limit any exfiltration attempts. So I’ve been fairly aggressive with the sandboxing.

* Is this just normal “noise” for a tightly sandboxed agent? The framework’s docs are vague on expected audit events.
* Or is it a sign my security policies are too restrictive and the agent is constantly fighting them, which might indicate a stability risk?
* How do you all triage what’s a real red flag vs. just expected friction in the sandboxing layer?

For comparison, when I ran a similar workload in Open Claw’s nano_claw with its default OCI sandbox, the audit log was much quieter. But I know the underlying isolation approaches are different.

Would love to hear your experiences, especially if you’ve run into similar log spam with other agent frameworks. What’s your threshold for investigating these ‘denied’ entries?