Anthropic’s high-performance AI model Mythos, which has shocked the global cybersecurity industry, was found to have identified over 10,000 high-risk and critical security vulnerabilities in core software worldwide within a month. While AI’s speed and accuracy in detecting security weaknesses have improved, human response times have relatively slowed, raising concerns that this is becoming a bottleneck in cybersecurity.
Anthropic disclosed the initial results of its AI security project ‘Gleamswing’ on the 22nd (local time). Using a preview version of Mythos, the company identified over 10,000 vulnerabilities through partner organizations participating in Gleamswing. For example, Cloudflare, an internet infrastructure service company, discovered 2,000 bugs in its core systems, of which 400 were high-risk or critical vulnerabilities. During the review process for a new version of Mozilla’s internet browser Firefox, 271 security vulnerabilities were found and fixed—over 10 times more than what previous models had detected. Some partner organizations also saw bug detection speeds improve by over 10 times.
External institutions also highly evaluated Mythos. The UK AI Safety Institute conducted a simulated test of Mythos’s hacking capabilities and reported that it successfully completed an attack process for the first time. The security platform XBOW also stated, “Mythos significantly outperformed existing models in web vulnerability attack performance evaluations and demonstrated unprecedented accuracy.”
Anthropic analyzed the results and noted, “The bottleneck in cybersecurity has shifted.” In the past, the focus was on how quickly new vulnerabilities could be found, but now, as AI identifies weaknesses too quickly and effectively, the slower human verification and problem-solving processes have become the bottleneck. Anthropic warned, “The time gap between vulnerability discovery and the deployment of fixes and patches will provide attackers with more opportunities, creating a new cybersecurity risk,” and advised, “Developers and companies must shorten their fix-and-patch cycles and deploy security updates more easily and quickly.”
Originally written by: Kang Da-eun
Image credit: Newsis
Source: The Chosun Daily
Published on: 25 May 2026
Link to original article: Anthropic’s Mythos AI Identifies 10,000 Vulnerabilities in Month
