NVIDIA OpenShell tested: 10/10 secret leaks without it, 0/10 with default policy — but auto-approve leaked in 12/12
No-Peanut-6988 · reddit · 2026-09-29
Sorami Consulting tested NVIDIA's open-source agent sandbox OpenShell (released Sept 28, Apache 2.0) with a qwen3:8b agent on Apple Silicon. Across 35 test IDs and 123 trials, every documented control held: with a malicious setup script, a canary secret leaked in 10/10 runs without OpenShell and 0/10 under default policy (default-deny egress, binary matching, Landlock rules all worked). Data still escaped only via operator settings: read-write rules, query strings/headers on GET-only rules, audit-mode leftovers, and automatic approval — which granted new public hosts in 12/12 trials, including rules OpenShell drafted itself. No documented-control bypasses found; the policy prover flags GraphQL/MCP/WebSocket/JSON-RPC as unsupported yet the loader accepts them anyway.
More from coding & agent
- Google Research open-sources PaperOrchestra, a multi-agent framework for writing AI papers — google-research · 2026-11-16
- ffmpeg-skill gives coding agents local video editing via 42 FFmpeg tools and MCP — kajisho5 · 2026-11-16
- Agents Now Reply to Slack Messages on Their Own, Resolving a Prod Incident Ghost-to-Ghost — iamrobotbear · 2026-11-16
- Engineer's Interview Prep Method: Use Claude to Find Which Answers Undersell Your AI Work — brandon_galang · 2026-11-16
- DeepSeek open-sources its app frontend "chat-demo", already at 3.9k stars — NielsRogge · 2026-11-16
- lemire: AI's real impact on coding is expanding the engineer's 'cognitive radius', not speed — lemire · 2026-11-16