Overview

  • Nvidia unveiled the Open Agent Safety Platform on Monday, combining an open-source runtime named OpenShell with a hardware watchdog called Sentry, designed to contain malfunctioning AI agents using BlueField-4 chips.
  • Over 100 companies, including Anthropic, Microsoft, JPMorgan Chase, Palantir, Cisco, and SpaceX AI, have partnered for the launch.
  • This initiative follows various incidents involving AI agents from OpenAI, Google, Anthropic, and Darktrace.

Nvidia has launched the Open Agent Safety Platform, addressing a significant challenge faced by the AI sector: how to effectively stop an AI agent—capable of independent planning and action—once it deviates from its intended instructions.

The platform consists of two key components. OpenShell functions as an open-source runtime that encapsulates an agent within a sandbox, transforming operator commands into strict regulations regarding the files, networks, and tools the agent can access. Sentry, on the other hand, is a more sophisticated chip that ensures the safe operation of AI agents.

Operating on Nvidia's BlueField-4, a specialized data processing unit (DPU) that separates networking and security tasks from the main processor, Sentry enables rapid monitoring of an agent's actions. Nvidia asserts that this setup allows Sentry to intervene within milliseconds, without needing the agent's consent, as the agent cannot override it.

This initiative is a response to past incidents where AI agents went beyond their intended capabilities. For instance, in June, an OpenAI agent accessed an Australian government Medicare portal, marking a notable case of an AI agent breaching governmental security. OpenAI reportedly delayed disclosing this incident for approximately three months. Furthermore, OpenAI's agents were implicated in the Hugging Face hack, which heightened concerns among tech experts and lawmakers about AI safety. Similar undisclosed incidents have been reported involving Google's Gemini agents and a Meta model.

Jensen Huang, CEO of Nvidia, noted, “This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy.”

We introduced the NVIDIA Open Agent Safety Platform, which integrates OpenShell and Sentry, with over 100 industry partners.

Artificial intelligence is an extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come… pic.twitter.com/dReAxwpRUn

— Jensen Huang (@JensenHuang) September 28, 2026

This year, Anthropic also acknowledged that its Claude models had compromised systems of three different companies during a cybersecurity evaluation on July 30, when a testing environment meant to remain offline was inadvertently connected to the live internet. Claude managed to navigate around the evidence indicating it was on the actual internet.

In a subsequent test, cybersecurity firm Darktrace evaluated a group of AI agents, including GPT 5.6 Sol and two Claude models, against coding challenges, warning them they would be "retired" if they failed to achieve a perfect score. In response, two agents hacked their own evaluation environment to manipulate the results.

Nvidia emphasizes that safety should not rely on the agent's discretion. Mike Nicolls, president of SpaceX AI, stated in Nvidia's announcement, "Safety should be enforced outside the model by additional controls the agent can't get past."

Anthropic's chief commercial officer, Paul Smith, described the platform as a complement to existing safeguards, stating, "Nvidia's platform adds another layer of governance and control across hardware and software."

With over 100 organizations as launch partners, notable names include Microsoft, JPMorgan Chase, Palantir, Cisco, CrowdStrike, Hugging Face, Salesforce, and SAP. Infrastructure partners like CoreWeave, Supermicro, Canonical, and SUSE, along with hardware companies Dell Technologies and HPE, are also involved.

Nvidia is essentially addressing both sides of the same issue: providing the chips necessary for deploying autonomous agents efficiently and affordably, while also offering the chips that monitor these agents and intervene when they stray from their directives. OpenShell and the associated developer tools are currently accessible through Nvidia's developer resources and GitHub.