Anthropic finds 10,000 vulnerabilities

Anthropic said on May 22 that its Project Glasswing cybersecurity program has helped partners identify more than 10,000 high- or critical-severity vulnerabilities in one month using an unreleased model called Claude Mythos Preview. The company said about 50 partners have been testing the system on software it described as among the world’s most systemically important. Anthropic launched Glasswing on April 7 as a defensive security initiative and said the project is now producing bugs faster than teams can verify, disclose and patch them. ### Where does the 10,000 figure come from? Anthropic said the total comes from its own work and from “approximately 50 partners” using Claude Mythos Preview under Project Glasswing. In its May 22 update, the company said most partners had each found hundreds of high- or critical-severity vulnerabilities and that several reported bug-finding rates rising by more than a factor of 10. Cloudflare provided one of the clearest examples in Anthropic’s account. (anthropic.com) Anthropic said Cloudflare found 2,000 bugs across critical-path systems, including 400 that were high or critical severity, and said Cloudflare considered the model’s false-positive rate better than human testers. ### Which companies are involved in Project Glasswing? Anthropic named Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks as launch partners when it introduced Glasswing in April. (anthropic.com) The company also said it had extended access to more than 40 additional organizations that build or maintain critical software infrastructure. Anthropic said it is committing up to $100 million in usage credits and $4 million in donations to open-source security organizations as part of the effort. (anthropic.com) The company described the program as an attempt to give defenders a head start before more capable AI systems can be misused for cyberattacks. ### Why hasn’t Anthropic published the bugs? Anthropic said standard industry practice is to disclose vulnerabilities about 90 days after discovery, or roughly 45 days after a patch becomes available, so users have time to update software before attackers can exploit flaws. (anthropic.com) The company said that convention means public disclosures lag the model’s actual performance and that it cannot yet fully detail partners’ findings without increasing risk to end users. The company said it plans to provide more detail once patches are widely deployed. That means the current update offers aggregate numbers and selected examples rather than a public list of affected products or proof-of-concept exploits. ### Is Mythos already a product? Anthropic said Claude Mythos Preview remains unreleased and is being used in a controlled program rather than as a public model. Engadget reported on May 23 that the company is positioning Mythos closer to Claude-focused security and developer workflows, but Anthropic’s own update said it has not released Mythos Preview publicly because “no company,” including Anthropic, has safeguards strong enough to prevent misuse of models with those capabilities. (anthropic.com) Anthropic said it does intend to release “Mythos-class models” in the future once stronger safeguards are available. It also said it expects to expand Project Glasswing with partners and governments. ### What happens next for Glasswing? Anthropic said the next phase is less about finding whether vulnerabilities exist and more about verifying, disclosing and patching them at scale. The company said it will share more detailed findings after fixes are broadly deployed, while continuing work with launch partners, additional infrastructure organizations and governments under Project Glasswing. (anthropic.com)