What unshare -n leaves behind: building a boundary for an AI recon agent
Why the boundary has to hold My AI agent's sandbox passed all 13 of its isolation tests on...
First seen by Oriel on .
AI summary
A GitHub Actions AI recon sandbox passed its isolation tests yet crossed its network boundary through a filesystem Unix socket, prompting a fresh-root jail and stronger layered controls.
- The original unshare -Urn jail blocked TCP and UDP but shared the host mount namespace, leaving reachable filesystem Unix sockets as channels to host-side processes.
- The replacement uses mount, network, PID, IPC and UTS namespaces, pivots into a minimal tmpfs root, exposes one broker socket, then drops privileges and capabilities.
- CI-driven hardening added loopback enforcement, a non-root broker, namespace-verified iptables and ip6tables owner rules, signed port limits, and a firewall launcher with broker supervision.
- Testing expanded from 13 isolation checks to whole-filesystem socket scans, negative controls, process limits, descriptor inventory, exact commit approvals, and archived SHA-256 evidence.
- The design remains a synthetic baseline: no independent attacker or live model has tested it, the broker lacks filesystem isolation, and cgroups and seccomp resource controls are absent.
#linux namespaces #sandboxing #github actions #network security
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Reactions
Total public reactions reported by the source; distinct from points and likes.
- DEV Community: 27, latest observation ; +10 over observed period since
Comments
Comment count reported by the source.
- DEV Community: 42, latest observation ; +13 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| DEV Community | 5000930 | https://dev.to/gnomeman4201/what-unshare-n-leaves-behind-building-a-boundary-for-an-ai-recon-agent-75b Read discussion |