Anthropic has introduced "Claude Mythos Preview," an AI model so proficient at discovering and weaponizing software vulnerabilities that its release has been restricted to a select group of 50 organizations under "Project Glasswing." The model demonstrated the ability to uncover thousands of zero-day bugs across major operating systems and browsers, including flaws unpatched for decades. This decision has sparked a debate on the balance between responsible disclosure and the lack of transparency regarding the model's false-positive rates and its potential to hallucinate vulnerabilities in corrected code.
The restriction of such powerful tools to a small circle of tech giants raises significant concerns about the security of systems outside the training data distribution, such as industrial control systems, medical firmware, and bespoke financial infrastructure. Critics argue that relying on the unilateral judgment of a private corporation to determine which pieces of global infrastructure are protected first creates a dangerous asymmetry. To mitigate these risks, experts call for globally coordinated frameworks for independent auditing and broader access for the academic research community to ensure collective security.
