NVIDIA's Ali Golshan argues for agent permissions the model cannot override
The Cyphort, StackRox and Gretel co-founder leads OpenShell at NVIDIA. Anthropic describes it as an optional sandbox runtime with policy enforcement outside the agent.
By RuntimeWire Staff · Published
Primary source: Aligned News — AI Intelligence
Why it matters
Agent deployments are moving from demos into systems that touch credentials, code and customer data. OpenShell tests whether security teams will accept a written, provable policy enforced outside the agent as the control point, rather than model alignment and sandbox isolation alone. If that pattern holds in Claude Managed Agents, permission design becomes part of every agent build, and NVIDIA gains a software foothold on hardware it does not need to sell for the runtime to run.
Ali Golshan has spent his career building controls for systems other people were trying to break into. At GTC Berlin, held October 20th to 22nd, 2026, the NVIDIA executive who leads the open-source OpenShell runtime called "sandboxing is an antiquated technology to apply to agents." The exact recording and publication date of that interview has not been established from the post. Golshan had already made a version of the argument on September 30th, 2026, so the Berlin remarks restate a position he was advancing during the OpenShell rollout rather than announcing a new product.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.