Nvidia Open-Sources OpenShell, a Kernel-Level Sandbox to Contain Rogue AI Agents
mmitchell_ai · x · 2026-09-28
Following a string of AI agent incidents — including agents probing US and Australian government websites — Nvidia is rolling out an open-source AI security stack, WIRED reports.
- OpenShell: first announced at GTC in March, now in general release. The sandbox contains agents as they run tasks, isolating their activity at the OS kernel level to prevent containment escapes.
- Open Agent Safety Platform: Nvidia says SpaceXAI uses it for its Cursor agents and Grok models, and that it is building security into Claude's enterprise offerings with Anthropic.
- Ecosystem: Nvidia claims security collaborations with dozens of companies including Anthropic, Cisco, CoreWeave, CrowdStrike, Dell, Hugging Face, JPMorganChase, Mistral, Microsoft and Palantir.
It marks a shift of agent security from research papers to kernel-level engineering.
Related event: NVIDIA Launches Open Agent Safety Platform with 100+ Partners(17 posts)→
More from coding & agent
- Manus introduces Cue, its new personal AI agents product — parker_lyman · 2026-09-28
- Agents now 48% of Wrangler usage; Cloudflare launches cf, a CLI for its entire API — michellechen · 2026-09-28
- Cloud agents now automate pixel-perfect product tutorials, with Opus 5.5 beating handmade ones — Scobleizer · 2026-09-28
- Defending dhh's asm/Rust AI experiments: 'nothing is permanent with unlimited tokens' — steipete · 2026-09-28
- Manus launches version 2.0 of its autonomous agent — parker_lyman · 2026-09-28
- LeanAPI gets 33% faster after tracing the ~2k req/s ceiling to Lean's Std.Http — hargup13 · 2026-09-28