Claude's Security AI Gets Sharper: Doubles Bug Finds, Tighter Sandbox

Anthropic has released new, specialized versions of its Claude AI model focused on cybersecurity. Claude Mythos 5.2, available only to vetted security organizations, and Claude Fable 5.2, its publicly available counterpart, share the same core technology but differ in their safety restrictions. According to Mozilla, which has been testing the models on Firefox code, Mythos 5.2 found roughly twice as many vulnerabilities as its predecessor in their scanning campaigns. Anthropic's internal benchmark showed Mythos 5.2 successfully identifying nearly 90% of known CVEs when given unpatched code.

For security engineers and bug hunters, this represents a significant step up in automated vulnerability discovery. The improved performance suggests AI tools are becoming more reliable as a 'second pair of eyes' for complex codebases, potentially catching subtle bugs other models miss. While access to the most powerful version is restricted, the public Fable model and the updated Claude Security scanner bring enhanced bug-finding capabilities to a wider audience, marking a tangible advance in AI-assisted security.