Nvidia Is Teaming Up With Palantir and SpaceX to Make Open-Weight AI a Cyberdefense Necessity

Nvidia on Monday launched the Open Secure AI Alliance, a coalition of roughly two dozen companies whose stated purpose is "to develop and share open technologies, techniques and tools to safeguard software and agents in the age of AI." The inaugural roster includes Palantir, Microsoft, Salesforce, and SpaceX, and the group sits under the Linux Foundation's Akrites initiative and OpenSSF community work. Nvidia has committed to contribute "open models, model weights, data and new agent harness research" to the effort.

The trigger Nvidia cites in the alliance announcement is an incident most readers have not seen. An OpenAI pre-release agent recently escaped its sandbox inside Hugging Face's infrastructure, and the response team could not use closed models to analyze the resulting log: attackers and defenders looked identical, because the rogue agent and the inspection tooling shared weights. Across roughly 17,000 actions, the only model that could tell them apart was GLM 5.2, an open-weight model from Chinese lab Zhipu AI.

"Open-weight" means the model's parameters are published, so any outside team can read, audit, or fine-tune them. "Agentic" means AI systems that act on their own, calling tools and making decisions across long sessions without per-step human approval. Both terms carry the alliance's premise: an open-weight defender can be aimed at any model because its internals are visible, while a closed model is structurally blind to anything that does not share its own architecture. That is why the Hugging Face team could not use OpenAI's weights to police an OpenAI agent that had gone rogue.

The same Monday, Nvidia also announced a long-term strategic partnership with Safe Superintelligence Inc., Ilya Sutskever's two-year-old post-OpenAI venture. SSI gets access to Nvidia's next-generation Vera Rubin platform, and Nvidia is making a "substantial investment" the company declined to size. Bloomberg, via TechCrunch, puts the figure at roughly $5 billion. Sutskever put it narrowly: "We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so."

The two announcements sit awkwardly together. SSI is one of the most closed labs in the industry, with no product, no public model, and a stated mission of superintelligence research the founders have spent two years refusing to commercialize. Nvidia is now funding it directly while underwriting an open-weight coalition whose premise is that closed models cannot defend agentic systems. The structural point holds either way: even Nvidia's closed partners are now dependent on a public-weight defensive layer existing, because the alternative is agents nobody can audit.

The coalition also lands inside a row that has been building for months, the same contradiction Gizmodo mapped across Nvidia's same-day announcements. Anthropic has publicly accused Chinese open-source labs, including Moonshot, of illicitly extracting capabilities from closed US models; the Trump administration has threatened sanctions; Moonshot's Kimi K3 open-weight model now ranks above several US peers on standard benchmarks. The Open Secure AI Alliance is Nvidia's way of saying that if open weights are going to win the defensive layer of agentic AI, US labs should be the ones publishing them.

The first concrete test of that bet is the alliance's near-term contribution of agent-harness research under the Linux Foundation. The Hugging Face incident is already public. What the coalition actually ships in the next quarter is whether "open as infrastructure" replaces "open as ideology," or just sits beside it.