Anthropic has introduced a powerful AI model called Mythos Preview, designed to identify critical software vulnerabilities while raising concerns about potential weaponization by malicious actors and nation-state hackers.

Key Points

• Anthropic is limiting access to Mythos Preview to 50 select organizations under a collaborative initiative titled Project Glasswing.
• The model has successfully identified high-severity vulnerabilities in major operating systems and web browsers, surpassing previous human-level bug-finding capabilities.
• Open-source maintainers, including those working on the Linux kernel and cURL, report that AI is significantly increasing the volume and accuracy of legitimate security vulnerability reports.
• Security experts warn that while AI aids defense, the future release of open-weight models could allow bad actors to generate exploit code without safety guardrails.
• The Pentagon has labeled Anthropic a "supply chain risk" due to the company's refusal to support autonomous weapons, a designation Anthropic is currently challenging in court.

Why it Matters

The rapid advancement of AI in cybersecurity creates a dual-use dilemma where the same tools used to patch critical internet infrastructure can be weaponized to disrupt global systems. As AI models increasingly outperform human researchers in finding bugs, the industry must balance the need for automated defense against the risk of powerful, unrestricted models falling into the hands of malicious actors.