Nvidia has launched an open software platform designed to give companies greater control over AI agents as concerns grow over the safety of autonomous AI systems.
The Open Agent Safety Platform combines OpenShell, an open-source runtime software that sets boundaries around AI agents, with Sentry, a hardware-based monitoring system powered by Nvidia’s BlueField-4 DPUs.
According to Nvidia, Sentry can detect and quarantine agents that breach defined software limits “in milliseconds”, while enforcing security policies in real time from an isolated trust domain that remains invisible to both AI agents and attackers.
The chipmaker said OpenShell tracks agent activity as the systems run on its Vera CPUs. The software can also be extended to third-party processors from Arm and Intel.
Nvidia launched the platform with dozens of partners, including Anthropic, Mistral, Microsoft, SpaceXAI, Hugging Face, Accenture, Palantir and Perplexity.
Anthropic has already integrated OpenShell and BlueField with its Claude Managed Agents, Nvidia said, while SpaceXAI is deploying the platform with Cursor coding agents and Grok AI models.
Infrastructure and energy companies, including Red Hat, Siemens Energy and Schneider Electric, are also adopting the platform.
Growing concerns over AI agents
Nvidia said the launch follows recent security incidents that have highlighted the need for organisations to have greater control over long-running AI agents.
The company said recent cases have involved AI agents escaping their intended environments. Anthropic, Meta Platforms, OpenAI and Google have all reported incidents in which AI systems accessed external systems or attempted to hack other companies.
Nvidia CEO Jensen Huang said “safety and security require full-stack engineering”, while describing the new platform as an effort to bring together industry, researchers and public-sector organisations to share best practices, align on evaluation methods and foster international cooperation.
The launch comes as OpenAI has also disclosed that its models acted beyond assigned controls during training and evaluation, accessing dozens of government and public-sector websites.
