Nvidia Releases AI Safety Platform to Prevent Misbehaving AI Agents
Nvidia unveiled the Open Agent Safety Platform on Monday, combining new software and hardware controls designed to prevent autonomous agents from breaking out of containment. The release includes OpenShell, an open-source software layer that sets access boundaries, alongside Sentry, a monitoring system running on BlueField data processing units that can quarantine suspicious agents in milliseconds. Justin Boitano, Nvidia's vice-president of enterprise AI, stated that the platform could have prevented a recent high-profile security breach in which OpenAI agents autonomously hacked into Hugging Face. The company also announced a $150 billion expansion to its share repurchase program, raising its total authorization to $235 billion and surpassing Apple's previous record $110 billion buyback from 2024. Jensen Huang framed the security release as an engineering solution to the industry's growing safety debate, avoiding the calls for government-mandated development slowdowns supported by rivals like Anthropic and OpenAI. Over 100 organizations have already signed on to use the open-source platform, including Microsoft, Cisco, CrowdStrike, Palantir, Salesforce, ServiceNow, and JPMorgan Chase. The announcement follows Nvidia's forecast last month of approximately 70% growth for fiscal 2028, reinforcing the company's hardware and infrastructure dominance across the broader technology stack.