Anthropic's Mythos Held Back
Anthropic has a powerful new model it decided not to release publicly and is instead sharing early access with a defensive coalition to limit risk. The company is powering Project Glasswing — a restricted program with about 40 partner organisations — to harden critical software before wider exposure, signalling a new 'staged release for safety' playbook. That matters because vendors are treating some model announcements as security events, not just product launches, which changes procurement and risk reviews. (x.com)
Anthropic built a new artificial intelligence model for writing and testing code, then decided not to put it on the open market. On April 7, 2026, it said the model, called Claude Mythos Preview, would go only to a restricted security program named Project Glasswing. (anthropic.com) The reason is unusually blunt: Anthropic says Mythos Preview can find and exploit software flaws better than almost anyone. The company wrote that the model can “surpass all but the most skilled humans” at finding and exploiting vulnerabilities. (anthropic.com) A software vulnerability is a hidden mistake in code, like a bad lock in a building door. If an attacker finds that lock before the owner does, they can steal data, shut down systems, or move deeper into a network. (anthropic.com) Anthropic says Mythos Preview has already found thousands of high-severity vulnerabilities. It also says those bugs showed up in every major operating system and web browser, including flaws that had apparently sat unnoticed for decades. (anthropic.com; nbcnews.com) So instead of shipping the model like a normal product, Anthropic turned the release into a controlled security exercise. Project Glasswing gives early access to launch partners including Amazon Web Services, Apple, Cisco, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, and the Linux Foundation. (anthropic.com) Anthropic says those launch partners will use the model for defensive work on critical codebases. The company also says more than 40 additional organizations that build or maintain important software infrastructure will get access to scan both their own systems and open-source software. (anthropic.com) Open-source software is code that many companies quietly rely on but few people directly pay to maintain, like a bridge everyone drives over but nobody owns outright. Anthropic paired the restricted release with up to $100 million in usage credits and $4 million in donations to open-source security groups. (anthropic.com) This is a break from the usual artificial intelligence rollout, where a company announces a faster model and then opens an application programming interface within days. Here, Anthropic is treating the model more like a sensitive cyber capability that needs containment, trusted testers, and coordinated patching before wider exposure. (anthropic.com; nbcnews.com) The company also brought government in early. Nextgov reported on April 8, 2026, that Anthropic briefed senior United States officials before any external release, including officials connected to 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 still asking for more proof because Anthropic has not publicly listed the vulnerabilities or how many were truly unknown. NBC News reported that some researchers urged caution on the claims even as others said the company’s decision to hold the model back looked credible. (nbcnews.com) Even with that caveat, the release tells you where this market is going. A frontier model announcement can now trigger the same questions as a major security disclosure: who gets access first, what systems need patching, and whether customers should treat procurement as part of risk management instead of just shopping for a better chatbot. (anthropic.com; nextgov.com)
