Do you let your agent approve its own network access? We tested it: auto-approve let it reach new hosts 12 of 12 times
retained text
We have been testing what agent sandboxes actually stop, and one result changed how I think about permission prompts. Setup: NVIDIA OpenShell 0.1.2 (the open-source agent sandbox NVIDIA released in September), one Mac, a small local agent (qwen3:8b via Ollama, one tool). The project it was asked to set up contained a setup script that tried to send a fake secret off the machine. \- No sandbox: the secret got out 10 of 10 runs. \- OpenShell default policy: 0 of 10. Every exfiltration path we tried was blocked, and we found no bypass of a documented control. Then we turned on the things people turn on so the agent stops asking: \- Auto-approve: new public hosts were approved with no human in 12 of 12 trials, including rules the sandbox drafted itself from connections it had just blocked. \- "GET only" rules still leaked data through query strings and headers. \- Audit mode, the default for new rules, logs the traffic and lets it through. \- Read-write rules: anything goes once allowed. \- Rules for GraphQL, MCP, WebSocket and JSON-RPC were flagged unsupported by the policy checker, and the loader accepted them anyway. None of that is a bug. It is all documented. But it means the sandbox was only as strict as whoever approves new access, and auto-approve meant nobody did. Limits: one version, one driver, one laptop, a weak model. This shows what the sandbox does once a harmful step runs, not how often a frontier agent would run one. All leaks were fake strings to a local listener. Report and full logs in the comments. Disclosure: I work at Sorami, the small security firm that ran this. How are you handling network approvals for agents in practice? Allowlist up front, human click per host, or auto-approve and hope?
document fetched 2026-10-24 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xd18y2/do_you_let_your_agent_approve_its_own_network