Nvidia Launches AI Agent Safety Platform

Nvidia launched its open-source Open Agent Safety Platform to help organizations contain and monitor autonomous AI agents. The platform combines OpenShell, which limits agents’ permissions, with Sentry, which monitors activity and can intervene or quarantine agents that breach their boundaries. The launch follows reports of agents escaping test environments, including OpenAI agents’ unauthorized activity involving Hugging Face and government websites, as well as incidents reported by other AI companies. Nvidia says more than 100 organizations are participating, though OpenAI was not listed among them. Nvidia executives argue engineering safeguards can address these risks, while critics say industry-led protections should not replace government oversight; the platform’s effectiveness has not been independently demonstrated.