NVIDIA's OpenShell Ditches Sandboxes for Zero-Privilege Agent Security
Hesamation · x · 2026-11-01
Hesamation interviewed Ali Golshan, who leads OpenShell, NVIDIA's open source agent runtime now built into Anthropic's Claude Managed Agents, at GTC Berlin. Golshan argues that "sandboxing is an antiquated technology to apply to agents."
OpenShell's core design differs from sandboxes:
- Agents start with zero privileges; access is granted only when necessary
- Secrets stay outside the containment, inaccessible to the agent directly
The conversation also covers whether AI security risks are overblown, whether new AI regulations are needed, whether OpenShell would have stopped the Hugging Face incident, alignment vs. security, and career advice for entering AI safety.
More from coding & agent
- Travel Agent MCP Server serves up-to-date visa and immigration data for trip planning — modelcontextprotocol · 2026-11-01
- Parse MCP connector reveals how any brand surfaces in ChatGPT and Google AI Overviews — modelcontextprotocol · 2026-11-01
- Dev claims Claude Opus 5.5 does in 1 day what 28 days of Unlimited Sol can't — karmay007 · 2026-11-01
- MCP creator leaves Anthropic after industry-wide adoption, stays on as core maintainer — bookwormengr · 2026-11-01
- Anthropic engineer: 90% of us build systems that prompt Opus 5.5 themselves — GregCook2011 · 2026-11-01
- Harness, Loops, Graphs: breaking down the 7 levels of agent autonomy, from babysitting to running a fleet — blaizedsouza · 2026-11-01