Claude Mythos constraints

Anthropic has a new model called Claude Mythos Preview, and the unusual part is not that it exists but that Anthropic is refusing to ship it like a normal chatbot update. On April 7, the company put it behind a restricted program called Project Glasswing instead of opening it to regular Claude users. (anthropic.com) The reason is cyber offense, not creative writing. Anthropic says Mythos Preview can find and exploit software vulnerabilities at a level beyond all but the most skilled human experts, which turns a language model into something closer to an automated bug hunter with lock-picking skills. (anthropic.com) A software vulnerability is a hidden mistake in code that can let an attacker break in, steal data, or shut systems down. Anthropic says Mythos Preview has already found thousands of high-severity vulnerabilities, including flaws in every major operating system and web browser it tested. (anthropic.com) That is why Anthropic did not start with consumers or even ordinary business customers. Project Glasswing launched with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, plus more than 40 other organizations that maintain critical software infrastructure. (anthropic.com) Anthropic says the model already surfaced bugs that had survived for years inside heavily used software. CyberScoop reported one example as a 27-year-old bug in OpenBSD and another as a 16-year-old flaw in FFmpeg that automated testing had missed even after millions of executions of the affected line. (cyberscoop.com) The company is spending real money to keep this defensive rollout narrow and fast. Anthropic committed up to $100 million in usage credits for Mythos Preview and another $4 million in direct donations to open-source security organizations tied to the Glasswing effort. (anthropic.com) The public first learned about Mythos in a much messier way. Fortune reported on March 26 that Anthropic had left roughly 3,000 unpublished assets in a publicly accessible content management system cache, and those draft materials described Mythos as a major jump in capability. (fortune.com) (tech.yahoo.com) That leak made the delay look, to some users, like a product launch that kept slipping. Anthropic’s actual April 7 announcement points the other way: the company says it does not plan to make Mythos Preview available to the general public and is limiting access because the same system that can patch critical code could also help attackers break it. (anthropic.com) (cyberscoop.com) The government piece is unusually explicit for an artificial intelligence model launch. Nextgov reported that Anthropic briefed senior United States officials before any external release, including officials at the Cybersecurity and Infrastructure Security Agency and the National Institute of Standards and Technology’s Center for AI Standards and Innovation. (nextgov.com) Outside experts are not treating this as marketing theater. NBC News quoted Luta Security chief executive Katie Moussouris saying the threat is real, while also noting that public evidence is still limited because Anthropic has not disclosed the full vulnerability list. (nbcnews.com) So the story is not “Anthropic built a better Claude and held it back for hype.” The story is that Anthropic appears to have built a model strong enough in codebreaking tasks that the first release path is a controlled cybersecurity pilot, with selected companies, government briefings, and no public launch date. (anthropic.com) (nextgov.com)