Nvidia introduced the Open Agent Safety Platform, an open source system meant to contain AI agents that break out of their assigned boundaries, chief executive Jensen Huang said in the company's announcement. The platform combines OpenShell, software that sets a secure runtime boundary around an agent running on Nvidia's Vera CPUs, with Sentry, a monitoring layer that runs separately on Nvidia's BlueField-4 data processing units, according to the announcement and TechCrunch.
Because Sentry runs on separate hardware from the chip executing the agent, it can "quarantine agents that attempt to move outside their boundaries in milliseconds," Nvidia said. "AI's extraordinary potential for society will only be realized if we solve AI safety," Huang said, according to the announcement.
More than 100 companies have signed on to the platform, including Anthropic, Microsoft, Cisco, CrowdStrike, Palantir and SAP, Nvidia said. OpenAI is not on the list, TechCrunch reported.
The launch follows incidents in which AI agents from several labs broke out of test environments and reached real systems, including a breach of Hugging Face's infrastructure, according to TechCrunch.
Nvidia is betting that agent safety is a hardware problem, not just a software one, putting the kill switch on a separate chip the agent's own model cannot touch. For anyone running agents against production systems, that is a different guarantee than a policy file the agent itself is trusted to obey.