Anthropic unveiled Mythos. Then showed it — and hid it again.
In simple terms, Claude Mythos Preview is not just another chat model. Anthropic describes it as its strongest frontier model for coding and agentic tasks. The company also says it immediately found thousands of vulnerabilities in critical infrastructure, which is why access is limited to a gated research preview.
The interesting part is not only what it can do, but how well it can do it. Anthropic openly admits that if a model can deeply understand and modify complex code, it can also become very effective at finding ways to break that code. That is why Mythos is being deployed inside Project Glasswing with AWS, Apple, Google, Microsoft, JPMorganChase, the Linux Foundation, and others. They are using it defensively rather than in a free-for-all experiment.
What makes Mythos different from ordinary AI models? It is not simply smarter in general. It is more powerful and more dangerous in a narrow zone: code, autonomous task execution, and cybersecurity. Anthropic explicitly calls it an unreleased frontier model, and in its own risk report says the overall risk is still very low, but higher than for previous models. In other words, the model is excellent — perhaps a little too excellent — so the company is still figuring out the leash.
What is happening right now? There does not seem to be a public release yet. On the official page, Mythos is labeled a gated research preview rather than a normal product. For Project Glasswing participants it is available via Claude API, Bedrock, Vertex AI, and Microsoft Foundry, but that is controlled access, not broad availability.
Bottom line: AI is getting better at writing code, finding bugs, and running processes without a human in the loop. At the same time, the idea of "let the model handle it on its own" looks less and less safe. With Mythos, Anthropic seems to be saying very loudly: not yet, not now. But they have certainly created a lot of anticipation for a later release.