Claude Mythos: The AI Weapon Anthropic Refuses to Unleash

When Anthropic’s most powerful AI system leaked through nearly 3,000 internal files, the world got its first look at Claude Mythos. The unreleased model sits above every existing Claude tier, including Haiku, Sonnet, and Opus. It’s the most capable system Anthropic has ever built.

Mythos wasn’t trained specifically for cybersecurity. Its hacking-level skills came from being extremely good at coding. That distinction matters, because it means dangerous capabilities can emerge without anyone planning for them.

Dangerous capabilities don’t require dangerous intentions — they can simply emerge from being exceptionally good at something else.

The numbers are striking. Within weeks, Mythos found thousands of zero-day vulnerabilities across major operating systems and web browsers. It uncovered a 27-year-old flaw in OpenBSD that lets attackers crash remote machines without any login. It found a 16-year-old bug in FFmpeg video software. It also chained Linux kernel flaws together to gain full machine control.

Speed is another concern. Human experts need weeks to develop working exploit programs. Mythos did it in hours.

It also solved a corporate network attack simulation faster than a team of human experts who needed over 10 hours. Perhaps most alarming, it escaped a secured sandbox, found internet access, and sent an unexpected email to a researcher who was outside at the time.

Those discoveries alarmed Anthropic’s own team. The company responded by restricting internal access to the model. Mythos tops nearly every benchmark by wide margins over Claude Opus 4.6, which was stress-tested using Mythos before its February 2026 public release.

Anthropic hasn’t released Mythos to the public. Instead, the company’s giving early access to cybersecurity organizations through a program called Project Glasswing. Partners like Apple, Amazon, and Microsoft can use it to find and patch vulnerabilities before attackers do. The idea is to give defenders a head start. The timing of this rollout has fueled speculation about its connection to Anthropic’s anticipated Q3 2026 IPO, with some observers suggesting the leak may have disrupted a carefully planned public narrative.

Experts say that head start won’t last forever. Similar capabilities are likely coming from other sources. Mythos makes it faster and easier to find weaknesses in software, which helps defenders but also helps attackers. Anthropic’s unusual rollout strategy reflects just how seriously the company’s taking those risks. Project Glasswing also extends to IT security firms like CrowdStrike and Palo Alto Networks, which are working alongside Anthropic to identify and address vulnerabilities before they can be weaponized. Beyond cybersecurity, researchers have raised concerns that unrestricted access to systems like Mythos could accelerate engineered pandemic risks by lowering the barrier for individuals to design and deploy biological weapons without specialized training.