Anthropic's Claude Mythos Is Too Powerful to Release

Anthropic just unveiled Claude Mythos Preview, an AI model so capable at finding software vulnerabilities that the company has decided not to release it to the public.

What Happened

During internal testing, Anthropic used Mythos Preview to scan major operating systems, web browsers, and widely-used open-source software. The model autonomously discovered thousands of previously unknown zero-day vulnerabilities — security flaws that the original developers never knew existed.

Some of these bugs had been hiding in plain sight for decades. The oldest was a 27-year-old vulnerability in OpenBSD. Another was a 17-year-old remote code execution flaw in FreeBSD that Mythos found, exploited, and documented entirely on its own. In testing, the model successfully reproduced and created proof-of-concept exploits on its first attempt in over 83% of cases.

Anthropic describes Mythos as a step change, not an incremental improvement, sitting in a brand-new performance tier above Claude Opus 4.6.

Why Anthropic Is Holding It Back

The same capabilities that make Mythos a powerful defensive tool also make it dangerous. If attackers had access to a model that could find and exploit vulnerabilities this effectively, the consequences could be severe. Anthropic says it is withholding public access until proper safeguards are in place.

Instead, the company launched Project Glasswing — a partnership program that gives select organisations early access to use Mythos for defensive security only. The initial partners include Amazon, Apple, Microsoft, CrowdStrike, Cisco, Broadcom, Palo Alto Networks, and the Linux Foundation.

What This Means for You

If you use Claude today, nothing changes immediately — Mythos Preview is not available through the regular Claude app or API. But the story matters for two reasons.

First, it signals that AI capabilities are advancing fast enough that even the companies building them are choosing restraint over speed to market. Second, it means the software you use every day — your browser, your operating system — is about to get a serious security audit from AI, which should ultimately make those tools safer.

For more on how Claude fits into your workflow today, see our complete Claude guide or check the latest on OpenAI’s competing GPT-5.4 models. If you’re just getting started with AI tools, head to our getting started page.