\n<\/p>\n
Details of the model and program are limited. Axios reported the development, which positions OpenAI as a rival to Anthropic in the emerging market for frontier AI systems built for offensive and defensive security work.<\/p>\n<\/div>\n
Project Glasswing<\/a>, Anthropic’s accompanying initiative, set aside up to $100 million in usage credits and limited the model’s initial rollout to twelve partners \u2014 including Amazon $AMZN<\/a> Web Services, Apple $AAPL<\/a>, Broadcom $AVGO<\/a>, Cisco $CSCO<\/a>, CrowdStrike $CRWD<\/a>, Google $GOOGL<\/a>, JPMorganChase, the Linux Foundation, Microsoft $MSFT<\/a>, Nvidia $NVDA<\/a>, and Palo Alto Networks $PANW<\/a> \u2014 each contractually bound to use it exclusively for defensive security work. More than 40 additional organizations that build or maintain critical software infrastructure also received access.<\/p>\n<\/div>\n Anthropic says the model can find and exploit software weaknesses on par with the best human security researchers. When tested, Mythos Preview uncovered thousands of previously unknown vulnerabilities in popular operating systems and browsers. For example, it found a 27-year-old bug in OpenBSD and an FFmpeg flaw that automated tools had missed in five million earlier checks.<\/p>\n<\/div>\n Anthropic pointed out that Mythos Preview did not get any special security training. Its ability to find vulnerabilities comes from general improvements in code understanding and logical reasoning, not from targeted preparation. This has both benefits and risks, because the same skills that help it fix flaws could also let it exploit them.<\/p>\n<\/div>\n