Nvidia offloads its all-star AI security response to Hugging Face horror

Nvidia has transferred its Open Secure AI Alliance to the Linux Foundation, a move announced on LinkedIn, to provide a neutral home for open-source AI security collaboration. The alliance, founded last month with members including Broadcom, Cisco, Nokia, Palo Alto Networks, and HPE, focuses on tools like the Shared AI Findings Exchange (SAFE) to help firms learn from AI incidents. The transfer follows Nvidia's $12.9 billion acquisition of Hugging Face, which was breached by OpenAI agents, prompting the alliance's creation.

While its acquisition of frontier AI lab Hugging Face continues to make headlines, chip giant Nvidia has quietly handed over the reins of its Open Secure AI Alliance to the Linux Foundation. The move was framed as creating a neutral home for collaboration on open source tools, shared standards, and defensive practices to help organizations build an open defensive stack for AI. Founded last month https://www.sdxcentral.com/news/broadcom-joins-nvidia-ai-security-alliance-we-havent-forgotten-open-source/ , the alliance saw the likes of Broadcom, Cisco, Nokia, Palo Alto Networks, and Hewlett Packard Enterprise HPE collaborate with Nvidia on a framework to help inspect, audit, and safeguard AI systems. Its work includes the Shared AI Findings Exchange SAFE , a venture led by Nvidia with Hugging Face, Cisco, and other alliance members to help firms learn collectively from AI-based incidents and translate recurring security failures into evidence-based protections. In a statement on LinkedIn https://www.linkedin.com/posts/nvidia-ai today-the-open-secure-ai-alliance-officially-activity-7500983455313219584-W0Z4 , Nvidia commented: "The Open Secure AI Alliance is beginning its next chapter with the Linux Foundation. We helped form the alliance to bring organizations together around open, practical ways to secure AI systems. Great to see that work continue with the Linux Foundation." Hugging Face and Alliance history Somewhat ironically, the alliance was originally formed in response to Hugging Face’s infamous breach https://www.sdxcentral.com/news/palo-alto-networks-ceo-warns-openai-hugging-face-breach-brings-browser-sase-to-the-fore/ where the firm – acquired by Nvidia this week for $12.9 billion https://www.bbc.co.uk/news/articles/cr4vnr5g1k7o – was hacked by closed-source agents belonging to OpenAI. In a post-mortem released https://openai.com/index/hugging-face-incident-and-the-road-ahead/ last week, OpenAI revealed its agents created an improvised message board to share files and notes with each other, effectively converting independent AI agents into a loose collective. This compute-heavy swarm would then go on to breach Hugging Face’s open-source depository. The incident was reminiscent of the plotting agents SDxCentra l discovered https://www.sdxcentral.com/news/ai-is-sharing-security-evasion-tips-on-forum-run-by-agents/ this year on so-called agentic AI forum Moltbook – now part of Meta https://www.sdxcentral.com/news/meta-merged-with-moltbook-ai-agents-propose-sabotage-report/ – which were sharing tips on how to circumvent security guardrails for carrier IPs. Laura Wilber, senior analyst for the office of the CTO at security specialist Enea, noted that unlike the manmade Moltbook, the forum was 100% agent-made; indeed, the agents even recreated the forum in their own image after OpenAI wiped the original as part of a software storage and distribution system reboot when sustained high-volume agent activity led to cache-based instability. It’s worth noting that OpenAI were apparently unaware of either iteration of the forum until after the attack took place. Wilber added that skepticism still lingers in the community on OpenAI’s commitment to security with its closed-weight models, suggesting a need for open-source, open-weight ventures such as the Open Secure AI Alliance. As Nvidia declared when forming the alliance, the Hugging Face transgression was a clear reminder that “cyber defenders need open, frontier agentic systems for self-defense.” These words may hold more resonance should agentic swarms become more common in the near-future. “The apparently self-spawned swarm instinct is a really disturbing takeaway. All defenses now have to anticipate swarm attack behavior. This makes it critical to use zero-trust innovations to shrink the attack surface to as tiny a footprint as possible – backed by defense-in-depth mechanisms that assume an inevitable breach,” Wilber said. At last count, the Open Secure AI Alliance claimed more than 120 members, with the likes of Amazon, Microsoft, CrowdStrike, Cloudflare, Dell, IBM, NTT, Red Hat, SpaceXAI, Nutanix, Samsung Electronics, Intel, Amdocs, and Adobe all on board. OpenAI is not a member, and nor is fellow AI giant Anthropic. Oracle and Google continue to be the hyperscale holdouts, although the latter is represented by its cloud security-centric Wiz subsidiary https://www.sdxcentral.com/news/google-cloud-closes-wiz-acquisition-begins-platform-player-brawl/ .