Anthropic said its new cybersecurity program, Project Glasswing, uncovered more than 10,000 software vulnerabilities during its first month. The initiative runs on Claude Mythos Preview, an unreleased AI model that the company says has significantly improved how quickly partners can identify bugs and broadened the range of serious issues they detect.
The report says partner organizations are now finding bugs at a rate more than ten times higher than before. Cloudflare identified 2,000 vulnerabilities, including 400 ranked as high or critical severity. Mozilla said it used the model to uncover and fix 271 Firefox vulnerabilities, a figure ten times higher than what it found in an older browser version using an earlier Claude model.
Anthropic said Microsoft’s recent warning that its patch releases will “continue trending larger for some time” is tied to the growing number of vulnerabilities uncovered through Mythos Preview.
Outside of partner collaborations, Anthropic said it used the model to scan 1,000 open-source projects over recent months, identifying 23,019 vulnerabilities, including 6,202 classified as high or critical severity. Separately, though not part of the formal report, a security research firm recently said it managed to breach macOS, long regarded as a secure operating system, by using Mythos’s bug-discovery capabilities.
Anthropic has not made Mythos Preview publicly available, saying that neither it nor any other organization has yet built safeguards strong enough to stop a model with these capabilities from being misused. The company says it plans to release what it describes as “Mythos-class models” once those protections are in place. In the meantime, Anthropic aims to broaden Project Glasswing through partnerships with the U.S. government and other governments, a move that could help mend ties with Washington.
The company is already teaming up with a list of big-name partners, including Amazon Web Services, Apple, CrowdStrike, Google, JPMorganChase, NVIDIA, and Palo Alto Networks, along with the partners mentioned in the vulnerability findings.
The report comes as Anthropic is reportedly nearing its first profitable quarter since launching in 2021. The Wall Street Journal recently reported that the company is on pace to generate $10.9 billion in revenue and $559 million in operating profit for the quarter ending in June. Anthropic, however, doesn’t expect profitability to continue as it plans to spend heavily on computing infrastructure and other costs.
Maybe you would like other interesting articles?