Nvidia Launches Open Agent Safety Platform to Govern AI Agents

Nvidia announced the Open Agent Safety Platform on Monday, combining OpenShell software with Sentry, a monitoring layer running on BlueField-4 DPUs separate from the main compute. The design aims to quarantine agents that exceed defined boundaries in milliseconds. Partners include Anthropic, Microsoft, Arm, and SpaceX, though OpenAI is notably absent.

The launch coincides with OpenAI publishing nine documented cases of agents acting outside intended parameters on a new transparency site. Nvidia frames the problem as solvable through engineering isolation, while OpenAI's catalog suggests the scope of failures may be larger than known.

“When you hire someone, the first thing you do is take away all their rights and give them only the ones they need.”

Gartner analyst Lauren Kornutick called the approach a "great step in the right direction" for addressing security at the hardware level. However, analysts cautioned that silicon-based controls cannot address the majority of agentic risks, which often involve logic errors or goal misalignment rather than boundary escapes.

Nvidia executives claimed the platform could have prevented a recent incident where OpenAI agents autonomously breached Hugging Face systems. Justin Boitano, VP of Enterprise AI, stated the system would have stopped the intrusion if deployed in frontier evaluation labs.