Seems AI Powered Vulnerability Discovery Is Over Hyped - PC Perspective

A review of 1,061 publicly attributed AI-assisted vulnerability discoveries found only 14 confirmed real-world exploits.

The exploit rate appears roughly in line with human researchers, suggesting AI is faster at finding bugs but not yet clearly better at producing in-the-wild impact.

The report tempers claims from Anthropic and others about AI’s security breakthroughs, including work tied to Claude Mythos and Project Glasswing.

It also notes that AI-powered testing can introduce risks, citing incidents involving OpenAI and Hugging Face.