Anthropic has published the first report on Project Glasswing, its cybersecurity initiative using Claude Mythos Preview. The unreleased model has helped partners discover over 10,000 vulnerabilities in just one month, with most partners finding hundreds of critical- or high-severity bugs. Cloudflare alone identified 2,000 vulnerabilities, while Mozilla found 271 in Firefox—10 times more than with previous models.

Mythos Preview increased partners' bug-finding rates by more than 10x. The model scanned 1,000 open-source projects, uncovering 6,202 high- and critical-severity vulnerabilities out of 23,019. Microsoft's recent larger patch releases directly resulted from Mythos findings. Despite these successes, Anthropic hasn't released Mythos publicly due to insufficient safeguards against potential misuse.

The cybersecurity initiative aligns with Anthropic's reported profitability for the first time since 2021, with $10.9 billion revenue and $559 million operating profit expected for the June quarter. The company plans to invest heavily in coming quarters while expanding Project Glasswing partnerships with governments including the US. Anthropic will release Mythos-class models only when adequate security safeguards become available.