Image Courtesy: Getty Images
AI company Anthropic says its unreleased cybersecurity-focused model, Claude Mythos Preview, has already helped uncover more than 10,000 software vulnerabilities just one month after launching its Project Glasswing initiative.
The company shared the early results in an initial report on Project Glasswing, a program introduced in April that uses AI systems to identify and prevent cyberattacks before they can be exploited. Anthropic says partner organizations using Mythos Preview have each discovered hundreds of critical and high-severity flaws in their software, with overall bug-finding rates increasing by more than ten times, according to Anthropic’s research report.
Among the companies participating in the initiative, Cloudflare reportedly identified around 2,000 vulnerabilities, including roughly 400 classified as critical or high severity. Mozilla also said it found and fixed 271 vulnerabilities in Firefox using Mythos Preview, a dramatic increase compared to earlier Claude-powered testing systems.
Anthropic says it also scanned 1,000 open-source software projects over recent months, identifying more than 6,200 high- and critical-severity vulnerabilities out of over 23,000 total findings. The results highlight how rapidly AI-assisted cybersecurity tools are evolving from experimental research projects into large-scale vulnerability discovery platforms.
The company has not publicly released Mythos Preview, citing concerns that models with advanced offensive cybersecurity capabilities could be abused if adequate safeguards are not in place. Anthropic says it intends to eventually release “Mythos-class” systems once stronger protections and monitoring frameworks are developed.
The report also signals growing alignment between major AI companies and government cybersecurity initiatives. Anthropic says it plans to expand Project Glasswing in collaboration with governments and enterprise partners, including Amazon Web Services, Apple, Google, NVIDIA, and JPMorganChase.
The findings arrive as software vendors face mounting pressure to secure increasingly complex systems against AI-assisted attacks. Major patch releases across the industry are becoming larger and more frequent as automated vulnerability discovery accelerates. Anthropic linked Microsoft’s recent comments about growing patch sizes to bugs identified through Mythos Preview.
The cybersecurity push also comes during a strong financial period for Anthropic. Reports suggest the company could become profitable for the first time since its founding in 2021, though executives reportedly expect future profits to fluctuate as spending on compute infrastructure and AI development continues to rise.
