Anthropic vs OpenAI: 271 bugs found, three of them real

The two largest A.I. labs both launched competing cybersecurity divisions this month. Anthropic's new Claude Mythos model was turned loose on the Firefox codebase. Per Mozilla, it identified two hundred seventy-one vulnerabilities. Mozilla called the model just as capable as an elite human researcher. Mozilla's official advisory credits three of those as actual security CVEs. A separate test was run on the curl codebase. The maintainer reviewed five reported issues. Three were already known. One wasn't a security bug at all. One was real, rated low. The curl maintainer Daniel Stenberg said the hype around the model was, in his words, primarily marketing. OpenAI's response, announced two days ago, is called Daybreak. Sam Altman said A.I. is already good, and about to get super good, at cybersecurity. Eight enterprise security firms have signed on as launch partners. Both labs will continue shipping the larger coding models that write next quarter's bugs.