Claude Mythos: The AI Challenging American Security

Claude Mythos: An AI with Remarkable Capabilities

Before its suspension due to a government directive, the artificial intelligence model Claude Mythos developed by Anthropic was used by select organizations to identify and rectify security vulnerabilities. Among these users, Mozilla, the creator of the Firefox browser, reported that this AI had enabled the discovery and repair of 217 security flaws in less than two months, vulnerabilities that had previously gone undetected. Claude Mythos demonstrated such effectiveness that it was able to identify flaws in some of the most secure software in the world, including those classified by the NSA.

Revelations from a U.S. Senator

These impressive capabilities were highlighted by Virginia Democratic Senator Mark Warner during a hearing before a Senate committee, as reported by Associated Press. Warner revealed that Claude Mythos had managed to detect breaches in nearly all classified systems in just a few hours. This information was corroborated by General Joshua Rudd, head of the NSA and the Cyber Command at the Pentagon. An anonymous source also confirmed that during a test, the AI had discovered vulnerabilities in sensitive and highly secure government computer systems. However, this source clarified that while the AI could identify flaws quickly, it did not mean it could exploit them just as fast. Before its access was suspended, Claude Mythos was offered by Anthropic for strictly defensive use.

Competition Intensifies with OpenAI

U.S. government testing took place before Anthropic decided to suspend Claude Mythos, along with its public version, Claude Fable 5. However, the race for innovation in cybersecurity does not stop there. This week, OpenAI unveiled its model GPT-5.5-Cyber, which surpasses Claude Mythos in performance on the CyberGym assessment, a test that measures AI models' ability to detect security flaws.

Towards the Generalization of Powerful AIs

Anthropic acknowledged that models similar to Claude Mythos could soon emerge, developed by entities potentially less concerned about security. "With the current pace of advancements in AI, these capabilities could quickly become widespread, perhaps even beyond responsible actors," the company warned in April. The implications of this generalization could be significant, affecting not only economies but also public and national security.

Claude Mythos demonstrated its power by detecting flaws in secure systems of the U.S. government within hours, as shown by tests conducted by the U.S. government. Although its access is currently suspended, it was previously offered to certain organizations for defensive use.