Anthropic, the company behind the Claude AI assistant, did something unusual with its Claude Mythos Preview model. It was strong enough that the company kept it out of general public use. Cybersecurity froze the rollout, because the model could find and exploit flaws in software at a level the company said surpassed all but the most skilled humans.
The official explanation begins with the bugs. In its Project Glasswing materials, the company said the model had found thousands of serious security flaws, including flaws in every major operating system and web browser. Under Glasswing, selected defenders got early access to repair critical systems, while broad availability waited until safeguards could contain the dangerous parts.
Independent testing supports that claim. The UK’s AI Security Institute tested the model in controlled conditions. It was the first system the institute tested to complete a simulated 32-step attack on a corporate network from start to finish. In a later evaluation, GPT-5.5, from ChatGPT maker OpenAI, reached a similar level. Restricting one model may slow one release, but it does not stop the same abilities from appearing elsewhere.
A second, more mundane explanation is computing power. The company has said demand for Claude grew sharply in 2026. Its agreement with Google and Broadcom, for multiple gigawatts of new computing capacity, begins in 2027. Neither fact proves a shortage held the model back, but both show the release decision sat inside a capacity crunch.
The third explanation is about rivals. A lab can train a weaker model on the answers of a stronger one, a practice called distillation. In February 2026, Anthropic said rival labs DeepSeek, Moonshot and MiniMax had used about 24,000 fraudulent accounts. Through them, the labs ran more than 16 million exchanges with Claude to extract its abilities. Safeguards on the later Mythos-class models treat distillation as a blocked use. Restricted access also slows rivals who collect answers to train competing systems.
Access then came in levels. Fable 5, a Mythos-class model with added safeguards, was meant for broad use. Preview users were to move up to Mythos 5, which has fewer restrictions. By June 12, 2026, a US government directive had led the company to switch off both models for all customers. Anthropic disputed the security basis for the order.
