Anthropic is accelerating the scale-up of Project Glasswing, onboarding approximately 150 new partners from 15 countries. The objective is ambitious: to automate the discovery of critical vulnerabilities in digital infrastructure that human lives depend on—energy, healthcare, and telecommunications. At the heart of this initiative is the Claude Mythos Preview model, specifically engineered to audit sectors where a potential cyberattack could impact over 100 million people. Even ENISA, the European Union Agency for Cybersecurity, is reportedly in talks to gain access to the platform.
Initial results from Glasswing read like a death sentence for traditional manual auditing. The first cohort of 50 participants has already identified more than 10,000 critical bugs. The speed and breadth of these specialized AI agents are transforming vulnerability research into an assembly line process that humans simply cannot compete with. However, Anthropic is playing a dual role here: both the savior and the herald of a digital apocalypse. The company warns that competitors could release similar models without stringent safety rails within six to twelve months, effectively handing out "master keys" for critical infrastructure to any interested party.
Behind this safety-first rhetoric lies a pragmatic business strategy. While Mythos Preview hunts for breaches, Anthropic is already aggressively marketing Claude Security, powered by Claude Opus 4.8—a tool designed to scan code and suggest patches. The strategy is flawlessly cynical: the company highlights a fire that its own technology could potentially ignite, then immediately offers exclusive extinguishers. Anthropic isn't just bolstering defenses; it is setting a precedent where the AI developer assumes responsibility for the security of software that its own models have learned to exploit so effectively.
Key Takeaways
Project Glasswing automates vulnerability discovery in critical infrastructure, including energy, telecom, and healthcare. The system has already identified 10,000 critical bugs, proving AI agents outperform manual security audits. Anthropic is simultaneously promoting its Claude Security tools, creating a closed loop of "detection and remediation."
"We are setting a precedent where the AI developer takes responsibility for the security of systems that its own models are capable of breaching."