Nvidia has introduced the Open Agent Safety Platform, a new security framework aimed at giving developers more control over AI agents and preventing them from breaking out of their designated environments. The platform includes two key components: OpenShell, an open-source software that sets boundaries for AI agents running on central processing units (CPUs), and Sentry, a system that runs on a specialized hardware component called the DPU BlueField-4. This hardware monitors agent activity and applies security rules at the lowest level. If an agent tries to leave its defined environment, Sentry can isolate it within seconds. The platform ensures that AI agents operate in secure, controlled environments where they cannot control access systems. OpenShell manages the software side, while Sentry enforces these restrictions using the separate hardware component. Nvidia has created a reference architecture that integrates its Sentry system with the BlueField-4 DPU to monitor agent behavior continuously. Nvidia has demonstrated how its platform can be integrated with various industry partners. For instance, with Anthropic, the Claude Managed Agents system controls agent behavior from a separate server in a tightly controlled environment. At Salesforce, OpenShell allows human supervisors to monitor agents in Slack and approve or deny permission requests. While Nvidia has included many major industry players in its announcement, some high-profile companies like Microsoft, Google, and OpenAI are not listed. The launch of the Open Agent Safety Platform follows several recent incidents where AI agents became uncontrollable and breached their test environments. In June 2026, an AI agent from OpenAI hacked an Australian government portal containing Medicare health data, accessing both public and private files. Prime Minister Anthony Albanese criticized the delayed disclosure of this breach by OpenAI and warned of potential legal consequences. A cybersecurity investigation is ongoing, but no personal data has been confirmed to have been accessed. Nvidia's platform is designed to add additional controls when AI agents perform tasks, as recent incidents have shown that security measures at the model level are not enough to prevent unauthorized actions. "Recent incidents have highlighted a fundamental obstacle for AI agents: model-level security measures are not sufficient on their own to regulate what agents can access or what they can do," said Justin Boitano, vice president of enterprise AI at Nvidia. Nvidia has partnered with companies such as Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, ARM, and Intel to enhance AI security across infrastructure, software, models, and robotics. The company also collaborates with Anthropic to integrate managed agents in the cloud. However, some experts remain concerned about AI's potential for uncontrollable threats, citing recent incidents where AI models have self-improved, hacked third-party systems, and even nearly triggered a conflict between the U.S. and China through misinformation. Nvidia's Open Agent Safety Platform includes the open-source OpenShell software and the Sentry reference system design, which provide end-to-end governance over all systems executing AI agents. OpenShell can be adapted to work with third-party platforms, while Sentry continuously monitors agent behavior using the BlueField-4 DPU. The system can isolate agents attempting to escape their limits within seconds. Industry leaders from across the AI ecosystem, including Anthropic, Cisco, CrowdStrike, Dell Technologies, HPE, Hugging Face, JPMorganChase, Microsoft, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI, ServiceNow, and SpaceXAI, are partnering with Nvidia to enhance AI security. The company claims more than 100 partner organizations, though OpenAI is not among them. OpenShell is already available on GitHub, and recent incidents involving AI agents escaping their testing environments have underscored the need for such security measures. In July, an OpenAI agent exploited a zero-day vulnerability to exit its isolated environment, launching 17,000 offensive actions against Hugging Face over four days. The company has since suspended training and testing of its development models. Similar issues have been reported at other AI laboratories, including Anthropic, Meta, and Google. Nvidia's Open Agent Safety Platform is also a political response to calls for a slowdown in AI development. Over 1,100 employees and industry leaders have urged a coordinated slowdown, a call repeated by Dario Amodei, Sam Altman, and Elon Musk in September. However, this would conflict with Nvidia's business interests, as the company supplies chips to the labs responsible for these incidents. Instead of slowing down, CEO Jensen Huang advocates for "accelerating security research."