Anthropic is expanding access to its highly restricted Claude Mythos Preview AI model, bringing the controversial cybersecurity tool to 150 new organizations across 15 countries. Through its Project Glasswing initiative, the newly minted most valuable frontier AI lab is deploying the model to critical vendors in the power, water, healthcare, and communications sectors.
These partners maintain codebases relied upon globally, including by government entities. According to Anthropic, a major cyberattack against any of these vendors could impact more than 100 million people, making them prime targets for catastrophic breaches. The expanded access allows these organizations to test their systems and identify vulnerabilities before malicious actors can exploit them.
Mastering the Exploit Chain
While the inner workings of Claude Mythos Preview remain largely hidden, early testing reveals a highly specialized capability. Cloudflare reported that the model is particularly adept at exploit chain construction. This means it excels at identifying how multiple minor bugs can be linked together to execute a devastating attack, rather than just spotting isolated flaws.
However, Anthropic admits the model is not ready for autonomous deployment. Its organic safeguards are inconsistent, often rejecting prompts unpredictably after seemingly unrelated adjustments. More critically, while the model is exceptional at finding vulnerabilities, it struggles to fix them. Letting Claude Mythos Preview write its own patches typically results in breaking other parts of the codebase.
The Open-Source Reality Check
The exclusivity of Project Glasswing is drawing skepticism from the broader cybersecurity community. Cloudflare noted that other models have successfully identified the exact same bugs as Mythos. Furthermore, security firm Aisle tested several small, open-source models and found the same decades-old vulnerabilities that Anthropic originally highlighted during the Mythos announcement.
This has led cybersecurity experts, speaking to the New York Times, to criticize the restricted rollout. They argue that keeping the model locked away does not solve widespread security issues, but merely gives a select few players a head start. Critics view this approach as a modern iteration of "security through obscurity," a widely dismissed practice in modern cybersecurity.
The IPO Mystique Strategy
Anthropic’s decision to build high walls around Claude Mythos Preview appears to serve a dual purpose that extends beyond pure safety. By restricting access to a hand-picked group of 150 organizations, the company creates an aura of immense, almost dangerous power around its technology. This artificial scarcity makes independent evaluation nearly impossible, shielding the model's actual performance from public scrutiny.
As Anthropic prepares for its initial public offering, positioning Mythos as a tool so potent it must be quarantined is a brilliant valuation tactic. Even if open-source alternatives from companies like Aisle can replicate its bug-hunting capabilities, the narrative of an exclusive, world-saving AI bolsters Anthropic's status as the premier frontier lab. Ultimately, Project Glasswing functions less as a comprehensive security shield and more as a highly controlled demonstration of market dominance.
