Severe security vulnerabilities surge around the launch of Claude Mythos Preview

Following the launch of Anthropic's Claude Mythos Preview, disclosures of high-risk and critical security vulnerabilities (CVEs) have increased sharply. In June 2026, about 1,500 high-risk and critical vulnerabilities were disclosed, more than 3.5 times higher than the monthly figures before the Mythos launch.

This phenomenon is related to Anthropic's announcement in April that Claude Mythos Preview can autonomously discover software vulnerabilities. Anthropic has been working with Microsoft, Google, Apple, AWS, and others through Project Glasswing to fix bugs before model releases, and is reported to have discovered over 10,000 vulnerabilities so far.

OpenAI is also making similar efforts toward defensive AI use through its Daybreak product. This is seen as an attempt to harden software for defensive purposes before frontier models can be used by malicious attackers.