Claude Mythos Unleashes Vulnerability Discovery Bottleneck

Anthropic's Claude Mythos has generated over 26,000 vulnerability findings since the launch of Project Glasswing in April 2026. However, only about 10% of these findings have reached the disclosure stage, with less than 1% actually patched. The analysis by security researcher Patrick Garrity indicates that human validation and coordination are becoming significant bottlenecks in the vulnerability remediation process. The findings challenge the narrative that AI can dramatically accelerate vulnerability discovery, as many generated flaws remain unverified. The situation has raised concerns about the effectiveness of the current vulnerability disclosure processes and the accuracy of Mythos's assessments. The urgency is compounded by the potential for bad actors to exploit these vulnerabilities before they are addressed. Governments are also alarmed about the implications of such powerful AI capabilities in cybersecurity.

Key Points: • Claude Mythos generated over 26,000 vulnerability findings since April 2026. • Only 10% of findings have reached the disclosure stage, with less than 1% patched. • Human validation is identified as a bottleneck in the vulnerability remediation process.