Anthropic's Project Glasswing finds 10,000+

Anthropic said on May 22 that its Project Glasswing cybersecurity program had already uncovered more than 10,000 high- or critical-severity vulnerabilities across widely used software, using an unreleased frontier model called Claude Mythos Preview. (anthropic.com) The company said it and about 50 partners had spent the month since Glasswing’s launch testing critical software before broader access to Mythos-class systems is considered. Anthropic said the findings came through a closed effort with companies and institutions including Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA and Palo Alto Networks. (red.anthropic.com) The announcement is one of the clearest public markers yet of how much pre-release security work advanced model developers say they are doing before putting more capable systems into wider hands. Anthropic said it is delaying general availability for Mythos-class models until stronger safeguards are in place to reduce risk across customers. ### What exactly did Glasswing find? Anthropic said the 10,000-plus total covers high- and critical-severity vulnerabilities in what it called “the most systemically important software in the world.” The company did not publish technical details for most of those flaws, saying more than 99% had not yet been patched and were being handled through coordinated vulnerability disclosure. (anthropic.com) Cloudflare, one of the participating organizations, said this month that Mythos Preview was able not just to spot suspected bugs but to write proof-of-concept code, test it in scratch environments and revise its approach when early attempts failed. (anthropic.com) Cloudflare said the model’s reasoning in some cases resembled that of a senior security researcher rather than a conventional automated scanner. ### Why is Anthropic keeping Mythos Preview closed? Anthropic said it does not plan to make Claude Mythos Preview generally available. In its technical materials, the company said the model could identify and exploit zero-day vulnerabilities in every major operating system and every major web browser when directed to do so, and could also turn known but unpatched flaws into working exploits. (anthropic.com) Opus 4.7, a separate Anthropic release announced last month, is being used as a stepping stone. Anthropic said that model includes safeguards designed to detect and block requests indicating prohibited or high-risk cybersecurity uses, and that lessons from its deployment will inform any eventual broad release of Mythos-class systems. (blog.cloudflare.com) ### Who is involved in the program? Anthropic said Project Glasswing launched with a group of large technology and infrastructure partners that includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. (red.anthropic.com) The company described the effort as a way to give defenders a head start on securing critical software before more capable offensive AI tools become more widely accessible. Around 50 partners are now participating, Anthropic said in its May 22 update. Interesting Engineering, citing the company’s announcement, reported that Anthropic is expanding the program further to additional partners and governments. (anthropic.com) ### Does 10,000 mean 10,000 confirmed public bugs? CSO Online reported in April that public confirmation lagged far behind Anthropic’s headline claims, citing analysis that found only one confirmed CVE directly tied to Glasswing at that point. That gap reflects the disclosure process as much as the underlying count: Anthropic has said the overwhelming majority of findings remain unpatched and cannot yet be publicly detailed. (anthropic.com) Red Hat said this week it was reviewing disclosed Glasswing findings for possible impact on products including RHEL and OpenShift, while adding that it had not identified confirmed significant vulnerabilities in its own environments so far. (anthropic.com) ### What happens before any wider release? Anthropic said its eventual goal is to enable users to deploy Mythos-class models safely at scale, but only after stronger safeguards are in place. The next visible milestones are likely to come through additional Glasswing updates, coordinated disclosures by affected software vendors, and Anthropic’s own model-release documents for future Mythos-class systems. (csoonline.com) (anthropic.com) (access.redhat.com)