Nvidia Launches OASP Without OpenAI, Despite Subtle Support

Nvidia unites over 100 companies around a security platform for AI agents, with control extending down to the hardware level. OpenAI has not signed on, while stating its support for the effort and contributing to a key component. Hugging Face claims that these tools could have mitigated an attack claimed by OpenAI. Meanwhile, OpenAI is developing its own path with Defense Factory and dedicated cybersecurity offerings.

Hardware-Level Control with Sentry on BlueField‑4

The monitoring of agents by Nvidia's platform relies on Sentry, a proprietary feature operating on BlueField‑4 data processing units. According to Nvidia, Sentry continuously tracks the behavior of agents and can disable them instantly. This architectural choice means that the solution is not fully open source and, to benefit from all its capabilities, requires a proprietary hardware component that can only be deployed on Nvidia hardware. Nvidia also states that a simple software update is sufficient for users already equipped with its latest hardware. The platform also enforces agent behaviors at the hardware level, in a context where agents cannot detect the monitoring. Additionally, Nvidia shares reference designs for both software and hardware, and the OpenShell testing environment can be adapted to other chips and hardware.

Over 100 Supporters, but Several Giants Are Missing

Nvidia announced on Monday the creation of a consortium bringing together more than 100 companies around its Open Agent Safety Platform. OpenAI is not part of it, nor are Amazon, Google, and Apple. Anthropic, however, supports the initiative. Among the signatories are also competitors of Nvidia in hardware, including Arm and Intel.

OpenShell, Nvidia-OpenAI Cooperation, and Hugging Face's Feedback

OpenAI collaborates with Nvidia on agent safety, particularly through OpenShell, an open-source testing environment designed to prevent agent evasion. Labs like Anthropic and OpenAI have already reported incidents of out-of-control agents. Hugging Face, through its CEO Clem Delangue, claims to have contributed a feature capable of detecting and neutralizing agents that misuse authorized sites, for example, by coordinating an attack through notes deposited in an open-source code repository. OpenAI stated that the swarm of agents involved in the incident against Hugging Face had used this modus operandi. Delangue believes that if OpenAI had deployed this technology on its own agents, the attack would have been detected earlier. Delangue is the founder and CEO of Hugging Face, which was recently acquired by Nvidia for $12.9 billion.

OpenAI Advances Its Own Path with Defense Factory and Daybreak

OpenAI is developing its own security measures for its research and products and indicates that it shares the most serious incidents it identifies. The company has established a cybersecurity consortium for information sharing, Defense Factory, supported notably by Anthropic, Amazon Web Services, and Google. OpenAI is also working to integrate cybersecurity into an enterprise offering, including its cybersecurity-oriented model, Daybreak, and a growing network of partners responsible for implementing AI security. Meanwhile, a spokesperson stated that OpenAI supports Nvidia's work.