We tested NVIDIA's new AI agent sandbox (OpenShell). The defaults held. Four settings let data out.
retained text
OpenShell is NVIDIA's open-source sandbox for AI agents: it blocks network, file and process access unless a policy allows it. We ran 123 trials on v0.1.2. The good: every documented control held, and a malicious script run by an agent leaked a secret 0/10 times under the default policy (10/10 without OpenShell). The four settings to check before you trust it: 1. Read-write rules let data out. 2. GET-only rules still carry data in query strings and headers. 3. Rules in audit mode log but do not block. 4. Auto-approval granted new public hosts with no human in 12/12 trials. Plain-language write-up and checklist: [https://sorami.com.au/guides/we-tested-nvidia-openshell/](https://sorami.com.au/guides/we-tested-nvidia-openshell/) Full report and logs: [https://sorami.com.au/research/nvidia-openshell-agent-sandbox-test/](https://sorami.com.au/research/nvidia-openshell-agent-sandbox-test/)
document fetched 2026-09-30 00:24 · canonical https://reddit.com/r/cybersecurity/comments/1wt5mnc/we_tested_nvidias_new_ai_agent_sandbox_openshell