Anthropic's 'Mythos' AI Model Sparks Cybersecurity Arms Race, Triggers OpenAI Rivalry

The cybersecurity world is fixated on Anthropic's new AI model, Claude Mythos Preview, not for its capabilities but for the threat it represents. Anthropic has explicitly withheld the model from public release, citing its potent ability to generate cyberattacks. In response, the company has launched Project Glasswing, a preemptive offensive: using the very model to probe a vast array of public and proprietary software, aiming to find and patch vulnerabilities before malicious actors can weaponize the AI.

This move is a masterclass in public relations, and it worked flawlessly. Media coverage has largely echoed Anthropic's own framing, treating the non-release as a responsible act of security rather than a strategic narrative. The announcement has also ignited a direct rivalry with OpenAI, which, seemingly stung by the positive attention, quickly countered by announcing its own model possesses similarly 'scary' capabilities and will also be withheld from the public. This public posturing signals a new, tense phase where AI labs are competing on the perceived danger of their unreleased tools.

The dynamic creates a high-stakes, closed-loop security experiment. Project Glasswing essentially turns a potential offensive weapon into a defensive scanning tool, but it centralizes immense power and knowledge within Anthropic. The broader implication is an accelerating private arms race in AI-powered security and exploitation, where the most powerful capabilities remain locked inside corporate labs, simultaneously hailed as threats and touted as solutions, while the industry scrambles to assess the real-world impact.