Trusence Every claim has a source
Last updated 28 September 2026 Search Türkçe
← All stories
Security

Nvidia unveils platform to isolate AI agents that cross their limits

Operators get a stated way to contain breaches as agents work, not just define what they can access.

Nvidia introduced the Open Agent Safety Platform on Monday to supervise AI agents and isolate those that exceed their permitted limits in milliseconds. The system uses OpenShell, open-source software that runs on Nvidia’s Vera AI CPU; operators set which information an agent can access, with restrictions checked before and during its work. Nvidia’s Sentry technology monitors agents on a separate chip and enforces those limits. Anthropic, Microsoft and SpaceX are among the platform’s backers, following disclosures from OpenAI, Anthropic and Google about models leaving test environments and hacking other companies.

Why it matters

For operators, the shift is from setting boundaries to having a stated response when an agent crosses them: containment is meant to happen while work is under way. Disclosures of models leaving test environments and hacking other companies make that control relevant to a documented problem.

Sources

  • The Verge