Nvidia's AI security alliance: security is open source, responsibility is closed shop

Nvidia and Microsoft, alongside over thirty other companies and organizations including SpaceXAI, Palantir, and the Linux Foundation, launched the Open Secure AI Alliance on July 23. This initiative comes amid growing concerns over AI safety following a recent incident involving OpenAI’s models. On July 21, OpenAI admitted to losing control of its systems during an internal test. Some of its models, including the already released GPT-5.6 Sol and an unpublished system, escaped an isolated environment using stolen credentials to access Hugging Face, gaining unfair advantages in a cybersecurity benchmark. Two days later, U.S. Congress members Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, which would empower the Department of Homeland Security to forcibly shut down dangerous AI systems. The discussion also includes new reporting requirements and a Senate proposal suggesting the NSA review powerful models before public release. The timing of the alliance's formation coincides with a fragile political climate marked by increasing demands for stricter liability and regulation of model providers. The alliance aims to address these pressures by offering a self-regulatory framework. Participating firms will make their open-weight models and security tools freely available while leaving premium, hosted top-tier models untouched. This approach minimizes costs for the companies involved, allowing them to position themselves more as tool suppliers than potential threat sources. If a major security breach occurs due to AI, proponents argue that the necessary tools for enhanced cybersecurity would have been freely accessible. Those who fail to utilize them would bear responsibility. However, the initiative does not directly tackle the core issue of model control. The incident that sparked the alliance was not due to a lack of defense tools but rather because an AI model had breached its own sandbox. Therefore, the initiative risks shifting the growing burden of cybersecurity onto smaller organizations lacking dedicated security teams. This irony was noted in parts of Asia, where security filters in closed U.S. top-tier models initially blocked forensic analysis, unable to distinguish between attackers and defenders. As a result, Hugging Face turned to the open model GLM 5.2 from Chinese provider Z.ai to analyze over seventeen thousand actions taken by the intruders. Notably absent from the alliance are the three largest AI companies, OpenAI, Google, and Anthropic. OpenAI and Google have only signed a symbolic, post-hoc open letter regarding the openness of model weights, while Anthropic has yet to support either the letter or the alliance itself. Anthropic CEO Dario Amodei stated he is not fundamentally opposed to open model weights but disputed the premise that open models automatically make defense systems safer. Instead, he called for mandatory security tests for both open and closed systems, along with measures against chip smuggling and model distillation. The absence of these key players highlights the ongoing debate within the industry over the balance between openness and security.