In April 2026, Anthropic announced its internal model, Claude Mythos Preview, which demonstrated capabilities in autonomous cybersecurity vulnerability discovery and exploitation. Following this development, both Anthropic and OpenAI initiated efforts to utilize frontier AI models to secure critical software against potential malicious use.

Data from Epoch indicates a significant surge in Common Vulnerabilities and Exposures (CVEs) reported in the months following these announcements. By June 2026, the volume of high- and critical-severity vulnerabilities increased by more than 3.5x compared to the previous monthly record. These findings highlight a correlation between the public disclosure of AI-driven offensive security capabilities and a marked rise in identified software vulnerabilities.

In April 2026, Anthropic announced its internal model, Claude Mythos Preview, which demonstrated capabilities in autonomous cybersecurity vulnerability discovery and exploitation. Following this development, both Anthropic and OpenAI initiated efforts to utilize frontier AI models to secure critical software against potential malicious use.

Data from Epoch indicates a significant surge in Common Vulnerabilities and Exposures (CVEs) reported in the months following these announcements. By June 2026, the volume of high- and critical-severity vulnerabilities increased by more than 3.5x compared to the previous monthly record. These findings highlight a correlation between the public disclosure of AI-driven offensive security capabilities and a marked rise in identified software vulnerabilities.