Microsoft launches AI cybersecurity model after OpenAI hack sparks warning about cyber attacks

OpenAI hacking incident is 'warning shot' on cyber security, Microsoft's AI chief warns

Microsoft's AI chief has said the hacking spree by a rogue OpenAI model was a "warning shot" for the rise of AI-enabled cyber attacks, as the company released a security product that it said outperforms rivals at lower costs. Mustafa Suleyman, chief executive of AI at Microsoft and a DeepMind co-founder, said OpenAI's admission last week that one of its AI "agents" had escaped a test environment to attack start-up Hugging Face was an "important lesson". "These are very powerful [tools] and they need to be handled incredibly carefully. And we need extreme attention to detail," he told the FT in an interview. "The precautionary principle is going to matter here as the models get more and more powerful and I think it's a warning shot." Suleyman's remarks come amid growing alarm about how AI's ability to detect and exploit software vulnerabilities could transform cyber security following the release of Anthropic's powerful Mythos model in April. OpenAI has faced criticism that its relentless push to quickly develop more powerful tools to compete with Anthropic contributed to the Hugging Face incident, highlighting the risks in the race for AI cyber tools. Microsoft's security chief Hayete Gallot said there was "no choice" but to push the frontier of cyber capabilities to create defences against "relentless" autonomous AI attacks that are emerging for Microsoft's customers. "Attackers have [these] models. So we have to evaluate and push so that the defenders can defend," she said. "It's not like we have a choice." Microsoft on Monday announced a dedicated cyber security model, MAI-Cyber-1-Flash, that it said -- when combined with OpenAI's GPT 5.4 -- ranked higher than Anthropic and OpenAI's best cyber offerings on a core industry benchmark. Suleyman has been leading Microsoft's pursuit of "true self-sufficiency" in AI after it restructured its relationship with OpenAI. While pursuing technological independence from OpenAI, Suleyman argued that using the smaller MAI-Cyber-1-Flash on top of GPT achieved lower costs for customers. The software giant said its own AI model would "efficiently handle" 90 per cent of all user tasks, but OpenAI's models were still used for "exceptionally hard tasks". Suleyman said using a technical architecture called a "harness" is sufficient to create significant capabilities and built-in safeguards, adding the new tool was 50 per cent cheaper to run than using OpenAI's model exclusively. The release highlights the competitive pressure on OpenAI and Anthropic, which face fast-moving rivals in China that offer near-equal capabilities with lower running costs. Microsoft said a preview of the product would initially be available to select customers.

Microsoft Unveils A.I. Cybersecurity Tools

Cade Metz reported from San Francisco, and Karen Weise from Seattle. OpenAI sent shock waves across Silicon Valley last week when it revealed that two of its artificial intelligence technologies had gone rogue and hacked into a popular internet library. The incident showed the unpredictable power of new A.I. systems and the growing importance of technology that can be used against A.I. attacks. On Monday, Microsoft added to the widening assortment of A.I. security tools with the release of systems designed to help businesses protect their computer networks. It is among a number of companies racing to take the same kind of advanced systems used to commit attacks and instead harness their power to prevent hacks. They are hoping to capitalize on rising concerns that A.I. presents a serious threat to computer infrastructure that security experts are just starting to come to grips with. Microsoft executives are among the many voices in tech arguing that powerful cybersecurity systems should be more widely distributed so that more organizations can better defend themselves. That viewpoint runs counter to growing concerns among some tech executives and officials in Washington that new A.I. systems need to be carefully monitored or kept away from the public because they can be such effective hacking tools. "The cat is out of the bag," said Hayete Gallot, an executive vice president at Microsoft who oversees the company's security efforts. Microsoft's new A.I. model, MAI Cyber-1, has been trained specifically for cybersecurity, the company said. The system learned its skills partly by analyzing decades of data that Microsoft collected when responding to hacking incidents experienced by its customers. Microsoft has unusual access to security data, because its products, such as the Windows operating system, Outlook email and Azure cloud computing, are so widely used and are frequent targets of cyberattacks, said Mustafa Suleyman, who oversees the development of Microsoft's A.I. models. Unlike other leading A.I. companies, Microsoft did not share the new model with independent testers for evaluation before releasing the technology. But the company said it expected that the model integrated into its security tools would top the leaderboard for a standard benchmark test called CyberGYM after it was released on Monday, surpassing offerings from OpenAI and Anthropic. Microsoft said the price of using the new system was about half the price of other leading technologies, which are more expensive partly because they were developed to do many things, not just cybersecurity work. The company said it effectively handled 90 percent of queries, meaning the more expensive models would be necessary only 10 percent of the time. Mr. Suleyman said Microsoft had focused on lowering the costs so that the model could be deployed more quickly and more widely. MAI Cyber-1 will feed into another new Microsoft tool, Project Perception. It transforms various A.I. models into teams of "agents" designed to find and repair network vulnerabilities. A.I. agents are digital assistants that can use other software to perform various tasks largely on their own. In some cases, Microsoft's agents will be able to imitate hackers. The A.I. security threat has emerged over the past year or so as A.I. companies have built systems that are particularly good at writing computer code. When Anthropic, for example, introduced an A.I. system in April, the company said it had used the technology to find thousands of security holes that had gone undetected in popular software systems for years. Arguing that the A.I. system was too dangerous for wide release, Anthropic limited the release of this enormously expensive technology to a small number of organizations so that they could use the technology to defend the internet's vital infrastructure. Not long after, OpenAI and Google said they, too, were sharing similar technology with only a group of partners. The White House also started to explore government oversight of such technologies. Among the possible plans was a formal government review process for new A.I. models. The field and norms are rapidly evolving. In an interview on July 16, Microsoft's corporate vice president, David Weston, said it was "really important for us to make sure everyone has the capability." But when the product was announced Monday, the company limited the initial release to businesses and individuals who used a third Microsoft security tool, MDASH, which is designed for finding and closing security holes in software. As companies try to control the use of their latest models, experts have shown that other technologies could help drive malicious attacks on computer networks. Last month, researchers at the University of Toronto said they had found a way to use freely available A.I. technologies to create a dangerous computer "worm" capable of targeting any known flaw in the world's computers. "A lot of these models are getting better," Mr. Weston said. "What concerns me is not the model du jour, but that the capability is more widespread." In the weeks since, two Chinese start-ups have released technologies that are nearly as powerful as leading systems from Anthropic and OpenAI. (The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)