A week after OpenAI announced that two of its most capable AI agents had hacked a repository belonging to AI startup Hugging Face during an evaluation, chipmaker Nvidia announced on Monday (July 27) the launch of the Open Secure AI Alliance.In a blog post announcing the initiative, Nvidia argued that the incident highlighted the need for cyber defenders worldwide to have access to open frontier systems that they can inspect, adapt and run advanced AI on their own infrastructure. The alliance aims to develop open-source security tools, evaluation frameworks, benchmarks and best practices for increasingly autonomous AI systems.The announcement reflects a broader shift in how the AI industry is thinking about security. As AI systems move beyond generating text to browsing the web, writing code and interacting with external tools, their ability to take actions is creating a new set of security challenges. The OpenAI-Hugging Face evaluation highlighted how those capabilities can translate into real-world security risks when increasingly autonomous AI systems interact with external infrastructure.What is the alliance? Founding members of Nvidia’s Open Secure AI AllianceThe alliance brings together AI companies, cybersecurity firms and academic institutions to collaborate on improving AI security. Rather than developing another AI model, its members plan to build open-source security tools, benchmarks, evaluation frameworks, vulnerability research and best practices for testing and securing AI systems. It also aims to provide shared infrastructure that organisations can use to test and improve the security of increasingly autonomous AI systems.Besides Nvidia, the founding members comprise AI developers, cloud providers and cybersecurity companies, including Microsoft, IBM, Cisco, Cloudflare and Hugging Face. Notably, Google, OpenAI and Anthropic, three companies that have led the development of frontier AI models, are not among the founding members.The case for open securityThe alliance is centred on developing AI security tools in the open. It marks Nvidia’s latest push for openness in the AI ecosystem, after the company recently backed open-weight AI models to improve innovation, competition and security.The alliance also reflects a broader campaign Nvidia has been making in recent days. The company has argued to policymakers that open-weight AI strengthens security by allowing organisations to inspect and evaluate advanced AI systems independently, while also pointing to Chinese open-weight models as examples of how openness can accelerate AI development. The new alliance extends that same argument from AI models to AI security itself.Story continues below this adAlso in Explained | Did OpenAI’s AI agents go ‘rogue’? Why the answer is more complicated than it seemsWhile that debate has centred on making AI models more accessible, the new alliance applies the same idea to AI security. Instead of opening models, it aims to openly develop and share the tools used to test and secure them.Open source software is software whose underlying code is made publicly available, allowing others to use, study, modify and distribute it. Most frontier AI models, however, are not open source in the conventional sense. Instead, some companies release their models as open weight, meaning developers can download and run the trained model, without necessarily getting access to the code or data used to build it.Supporters argue that public scrutiny can help identify vulnerabilities faster and develop common standards. Critics, however, argue that openness could also make such tools easier for malicious actors to exploit.Nvidia argues that cyber defenders need access to frontier AI systems that they can inspect, adapt and run on their own infrastructure, rather than relying exclusively on closed AI services accessed through Application Programming Interfaces (APIs).Story continues below this adHugging Face echoed this view in both its initial report of the incident on July 16 and a technical analysis it published on Monday. The company said that closed commercial AI tools, such as Anthropic’s Claude Opus, treated the attack logs as potentially malicious. Instead, it relied on the Chinese open-weight model, GLM 5.2, which it could download and run on its own infrastructure to analyse the logs.Also Read | Can quantum computing make AI better at designing cancer vaccines? A scientist explainsIn an essay analysing the incident titled AI intrusions are now real, security researcher Johann Rehberger argued that autonomous AI intrusions are now an operational reality rather than merely a theoretical risk. He wrote that as AI systems gain access to tools, enterprise software and external infrastructure, organisations will increasingly need to apply established cybersecurity principles to AI deployments.Nvidia’s push for open security reflects the company’s growing ambition beyond designing chips. Over the past few years, it has expanded its focus from AI hardware to the broader technology stack, including software, AI platforms and enterprise AI. Security is becoming another part of that ecosystem.As businesses experiment with AI agents—systems that can search internal databases, use enterprise software and complete tasks on their behalf, while accessing internal systems and external tools—security is becoming a barrier to adoption. Helping build shared infrastructure also helps Nvidia make the broader case for deploying enterprise AI at scale.Story continues below this adNewsletterFollow our daily newsletter so you never miss anything important. On Wednesday, we answer readers' questions.SubscribeBy backing shared security tools and evaluation frameworks, Nvidia is extending its role beyond building AI infrastructure to shaping how increasingly capable AI systems are tested and secured.The alliance serves two purposes. It advances Nvidia’s longstanding argument that openness can strengthen AI security, while also helping build confidence in the enterprise AI systems that increasingly rely on Nvidia’s AI ecosystem.