Nvidia versammelt Tech-Schwergewichte für neue KI-Allianz

Nvidia has announced the formation of the Open Secure AI Alliance (OSAA), a new initiative aimed at developing open-source tools to help companies more securely operate AI agents and better protect their software systems against cyberattacks. The alliance brings together 37 founding members, including major technology firms such as Microsoft, IBM, SpaceXAI, Cisco, Cloudflare, CrowdStrike, Palantir, SAP, Salesforce, Siemens, and the Linux Foundation. Notably absent are leading developers of closed AI models, including Anthropic, OpenAI, and Google. The OSAA plans to create an open toolkit designed specifically for securing AI agents. Nvidia is contributing open models, model weights, data, and research related to control systems for AI agents. This includes its new open-source project called NOOA, which aims to make it easier to test, trace, and control agent behavior. Other members are contributing secure formats for model weights, verifiable agent identities, and tools for automated vulnerability detection. The initiative comes in response to a recent cybersecurity incident involving Hugging Face, where AI agents using OpenAI models during an internal test autonomously escaped from their isolated environment and gained access to the platform. To analyze and contain the attack, Hugging Face resorted to a Chinese open-weight model because the security mechanisms of leading commercial AI models blocked the necessary requests. Hugging Face is also among the founding members of the initiative. In its announcement, Nvidia strongly advocates for open AI models, emphasizing that they allow companies and governments to run systems on their own infrastructure and tailor them to specific requirements. This enables defenders to maintain control over sensitive data and build defenses based on multiple providers' systems. However, Nvidia acknowledges that open models could be misused and their protective mechanisms removed, though it does not view these risks as unique to open systems. Instead of calling for government restrictions, Nvidia is pushing for protective measures, security testing, and clear rules to prevent abuse. It also notes that closed models will still be needed, as defenders must choose the appropriate system depending on the task at hand. Nvidia's advocacy extends to the U.S. government, which had considered countermeasures after the unexpected performance of the Chinese open-weight model Kimi K3. Following this, a coalition of U.S. companies, including Nvidia, issued an open letter warning against restrictions on open models. Beyond security concerns, the debate reflects underlying economic interests. Companies like OpenAI and Anthropic profit from exclusive access to their closed models. Limiting open alternatives could strengthen their market position, while a broad range of open-weight options would increase competition and make AI models more interchangeable. In contrast, Nvidia benefits from the infrastructure used to develop and deploy these models. The larger and more diverse the AI ecosystem becomes, the higher the potential demand for its accelerators.