Cybersecurity Tool

• Anthropic is investigating unauthorized access to its Mythos AI tool, a critical cybersecurity system, after reports revealed that an unauthorized group breached the platform via a vendor vulnerability. • The incident raises fresh concerns about security gaps within advanced AI systems and amplifies questions about the trustworthiness of AI tools handling sensitive security functions. • The breach exemplifies broader risks in the interconnected tech ecosystem, where vendor compromises can cascade into exposures of high-value AI infrastructure.

• The U.S. government is evaluating a restricted rollout of Anthropic's Mythos frontier AI model to federal agencies under Project Glasswing for defensive cybersecurity purposes. • Mythos has identified thousands of vulnerabilities across operating systems and web infrastructure at unprecedented speed, far surpassing traditional manual audits that take months or years. • Officials emphasize collaboration with model providers and intelligence community to implement guardrails before wider agency access, as stated by spokesperson Barbaccia.

• Anthropic discovered its AI tool can outperform humans at certain hacking and cybersecurity tasks, marking a significant milestone in autonomous security capabilities. • The findings have prompted discussions among regulators and legislators regarding the implications of advanced AI systems in cybersecurity applications. • The development highlights growing concerns about balancing AI capabilities with security risks, particularly as AI systems demonstrate competency in traditionally human-dominated security roles.

• Anthropic and OpenAI are both developing advanced AI cybersecurity products in a competitive race to control offensive and defensive capabilities, with OpenAI finalizing a security product for limited partner release and Anthropic running Project Glasswing internally. • Project Glasswing brings together major tech companies including Amazon, Apple, Microsoft, Google, and others to proactively hunt critical software vulnerabilities before attackers exploit them. • A joint study by Anthropic and MATS Fellows found that Claude Sonnet and GPT-5 could produce exploits against Ethereum smart contracts worth $4.6 million and uncovered two novel zero-day vulnerabilities in nearly 3,000 recently deployed contracts.