Anthropic shelves Mythos

Anthropic has not released its new Mythos model to the public, limiting it to a small research program after internal safety testing flagged cyber risk. (anthropic.com) Anthropic published a system card for “Claude Mythos Preview” on April 7 and called it its “most capable frontier model to date,” with stronger scores than Claude Opus 4.6 on coding and other benchmarks. The same document says the company released it only to “a small set of external customers” after first deploying it internally. (anthropic.com, anthropic.com) The company’s security research group said Mythos Preview could identify and exploit previously unknown software flaws in “every major operating system and every major web browser” during testing. Anthropic said those findings led it to keep the model out of general release. (anthropic.com, anthropic.com) That decision lands in the middle of a broader fight over how frontier artificial intelligence models should be deployed when they can write code, automate tasks, and search large codebases faster than human teams. Anthropic said Mythos is a general-purpose model whose cybersecurity strength comes from the same software-understanding ability that makes it useful for engineering work. (anthropic.com, anthropic.com) Anthropic paired the limited release with “Project Glasswing,” a program announced this week to give Mythos access to launch partners and more than 40 additional organizations that maintain critical software infrastructure. The company said those groups will use the model for defensive scanning and hardening work, not open public access. (anthropic.com) Semafor reported on April 8 that Anthropic told partners Mythos had exposed thousands of vulnerabilities in widely used software that did not yet have fixes. TechCrunch reported on April 9 that Anthropic’s rollout also drew skepticism from some industry figures who argued the company had an incentive to reserve top models for enterprise customers. (semafor.com, techcrunch.com) Anthropic’s separate alignment risk report reached a narrower conclusion on a different question: it said Mythos appeared to be the “best-aligned model” the company had released, even as it warned the model was more capable and more agentic than earlier systems. In other words, the company’s public concern was not that Mythos would independently go rogue, but that people could use its coding and security skills in dangerous ways. (anthropic.com) Business Insider reported on April 11 that some cybersecurity experts think the danger is real while others think Anthropic may be overstating how much one model changes an already bad security landscape. Anthropic has not announced a date for a broader Mythos launch. (businessinsider.com, anthropic.com)