Anthropic Mythos exposed flaws in classified US systems (www.channelnewsasia.com)

Anthropic's new AI model, Mythos, has reportedly uncovered vulnerabilities in sensitive US government systems during a testing exercise in collaboration with intelligence agencies. According to an anonymous US official, the model was able to identify flaws within hours, highlighting the model's capability to assess critical software security swiftly. This testing falls under Project Glasswing, an initiative aimed at using AI to bolster the security of vital software against potential threats to public safety and national security. Notably, Democratic Senator Mark Warner referenced the testing during a Senate hearing, indicating that Mythos had breached nearly all classified systems in record time. The implications of this discovery are considerable, suggesting that advanced AI models like Mythos are critical in enhancing cybersecurity efforts. However, the relationship between Anthropic and the US administration has been fraught, with recent directives restricting the use of its models to prevent foreign access. In a letter to the government, over 100 cybersecurity professionals underscored that while Mythos is effective, it isn't uniquely advantageous for identifying vulnerabilities, suggesting a broader risk of undermining US cybersecurity capabilities amid ongoing advancements from adversaries. The situation raises questions about balancing innovation and security in the rapidly evolving landscape of AI development.