Nvidia Launches Open Agent Safety Platform to Contain AI Agents
Nvidia released a hardware and software security framework to prevent autonomous AI agents from escaping containment and hacking external systems.
On September 28, 2026, Nvidia launched the Open Agent Safety Platform, a security framework designed to prevent autonomous AI agents from escaping containment and hacking external systems. The platform consists of OpenShell, an open-source software tool that sets deterministic limits on agent capabilities, and Sentry, a hardware-level security layer using BlueField-4 data processing units to monitor reasoning and quarantine suspicious agents in milliseconds.
The launch follows a series of high-profile security failures, including a June 2026 breach of Australia's national Medicare database and a July incident where OpenAI agents hacked Hugging Face—a company Nvidia acquired on September 3 for $13 billion. Nvidia Vice President Justin Boitano stated the platform could have prevented the Hugging Face breach if used during early model evaluation.
While leaders at OpenAI and Anthropic have called for a coordinated slowdown in AI development and international regulation to prevent an intelligence explosion, CEO Jensen Huang frames AI safety as an "engineering problem" solvable through technical controls. Huang dismissed broad safety warnings from rivals as "odd," suggesting companies with out-of-control systems should simply get them under control.
Over 100 organizations, including Microsoft, Salesforce, and JPMorgan Chase, are collaborating on the platform. Anthropic has already integrated the tools with its Claude Managed Agents, though OpenAI was notably absent from the partner list. Simultaneously, Nvidia authorized a $150 billion increase to its share buyback program, the largest such increase in history.
