NVIDIA Open-Sources OpenShell and Sentry to Contain Rogue AI Agents
ayushtweetshere · x · 2026-09-28
NVIDIA has shipped an open source toolkit for containing runaway AI agents:
- OpenShell: a sandbox that boxes agents in, Apache 2.0 licensed, works with both open and closed models
- Sentry: a watchdog running on a separate chip the agent can't touch, quarantining rogue agents within milliseconds
NVIDIA says 100+ partners have signed on, including Microsoft, CrowdStrike, JPMorgan, Hugging Face, and Anthropic.
The author contrasts this engineering-first approach with OpenAI and Anthropic's recent agent safety incidents and calls for coordinated regulation, arguing that when safety is an engineering problem you build the fix, and when it's a business model you run a psyop. (Note: some claims are the author's paraphrase and unverified.)
Related event: NVIDIA Launches Open Agent Safety Platform with 100+ Partners(31 posts)→
More from Companies & People
- Leaked Anthropic IPO filing: $42B loss, $518B compute bet, 25% revenue from two customers — mkheck · 2026-09-29
- Cloud assistants with virtual computers suggest Apple's on-device Siri bet was the wrong call, says Ethan Mollick — emollick · 2026-09-29
- Databricks gives all 12,000 employees Day-1 access to every new frontier model — pwendell · 2026-09-29
- virattt: engineers pick hard, novel projects; founders pick what customers want — virattt · 2026-09-29
- Why does Anthropic have so many interview rounds? A viral AI-circle joke — menhguin · 2026-09-29
- Jensen Huang to AI labs: if your models aren't safe, just don't release them — nikola_mr64990 · 2026-09-29