AI BriefingNVIDIA OpenShell Sandbox Blocks Leaks in Default Mode, But Convenience Settings Allow Data Exfiltration
·2026.10.23 10:40
Key point
Tests show NVIDIA OpenShell blocks malicious scripts in default mode (0/10 leaks), but enabling auto-approve allows data exfiltration in 12 of 12 trials.
Details
- NVIDIA OpenShell, a free open-source agent sandbox, prevented a malicious setup script from leaking secrets in 10 out of 10 tests when using default settings.
- Enabling the auto-approve convenience feature allowed data exfiltration in 12 of 12 trials, including cases where the sandbox wrote its own rules for newly approved connections.
- Other convenience settings, such as "read-only" rules and audit mode (the default for new rules), also permitted data leakage via web addresses, request headers, or logged traffic, though specific trial counts for these individual settings were not provided in the text.
- The study used a small local model on a single Mac, noting that results reflect sandbox behavior rather than the decision-making frequency of more advanced agents.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Nvidia OpenShell: Secure Runtime for Autonomous AI Agent Fleets