Anthropic Project Glasswing finds 10,000 vulnerabilities

Anthropic said on May 22 that its Project Glasswing cybersecurity program had found more than 10,000 high- or critical-severity vulnerabilities in roughly its first month, using the company’s Claude Mythos Preview model. The company said the work was carried out with about 50 partners scanning “the most systemically important software in the world,” with many of the flaws described as previously unknown to developers. Anthropic has not released Mythos Preview publicly and said the model is being restricted because of its cyber capabilities. HotHardware reported the update on May 23, citing Anthropic’s initial public accounting of the program. ### Where did the 10,000-vulnerability figure come from? Anthropic’s May 22 update said “we and our approximately 50 partners” used Claude Mythos Preview to find more than 10,000 high- or critical-severity vulnerabilities across critical software systems. The company said the figure covered findings generated since the April launch of Project Glasswing. Anthropic’s April 7 launch page said the initiative started with launch partners including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, and that access was later extended to more than 40 additional organizations. (anthropic.com) Anthropic’s public materials describe the flaws as including zero-day vulnerabilities, meaning bugs that were previously unknown to the software’s developers. On its Glasswing page, the company said Mythos Preview had identified “thousands of zero-day vulnerabilities,” many of them critical, across every major operating system and every major web browser, along with other important software. ### What exactly is Claude Mythos Preview doing? (anthropic.com) Claude Mythos Preview is Anthropic’s unreleased frontier model for cybersecurity work, according to the company’s research and product pages. Anthropic said the model can identify and fix vulnerabilities across hardware and software, and in some cases can reverse-engineer exploits on closed-source software and turn known-but-unpatched vulnerabilities into exploits. (anthropic.com) Anthropic said the model is “strikingly capable” at computer security tasks and that it does not plan to make Mythos Preview generally available. The company said its goal is to let users deploy Mythos-class models safely at scale later, but for now it is limiting access through projects such as Glasswing. ### Why isn’t Anthropic naming the bugs? Anthropic said more than 99% of the vulnerabilities it has found have not yet been patched, which is why it is withholding technical details. (red.anthropic.com) The company said disclosure is being handled through a coordinated vulnerability disclosure process rather than public release of exploit information. The May 22 update said the pace of discovery is outrunning remediation in some cases. (red.anthropic.com) External reports citing Anthropic’s update said only a fraction of confirmed flaws had been patched so far, though Anthropic’s own public post emphasized the disclosure process and did not publish a full bug-by-bug tally. ### Which systems are affected? Anthropic’s public statements say the findings span every major operating system and every major web browser, plus other widely used software. (red.anthropic.com) The company has not published a full affected-software list, but it has characterized the targets as software with broad systemic importance. The April 7 launch announcement said Anthropic was committing up to $100 million in usage credits and $4 million in donations to open-source security organizations to support the work. (anthropic.com) That same announcement framed the effort as preparation for a period when similar cyber capabilities could spread beyond companies focused on defensive use. ### What happens next? Project Glasswing remains active as of Anthropic’s May 22 update, with partner organizations continuing to use Claude Mythos Preview in defensive security work. (anthropic.com) Anthropic’s newsroom lists the May 22 “initial update” as the latest public milestone for the program, and the company has said it will share what it learns so the broader industry can benefit. Anthropic’s next public disclosures are likely to come through additional Glasswing updates or partner-specific vulnerability advisories as patches are completed. (anthropic.com) For now, the named participants remain the companies and institutions listed at launch, including Apple, Google, Microsoft, AWS, Cisco and the Linux Foundation. (anthropic.com)