Anthropic teases 'Mythos' model
Anthropic revealed a powerful new model called 'Mythos' and says it's delaying a public release while also working on Project Glasswing, signalling caution even as capabilities advance. The company’s choice to hold back wider access suggests a focus on controlled rollouts or additional safety and capability work before general availability. That posture matters because it can slow competitor pressure but also fuels speculation about capability gaps and release timing. (x.com) (x.com)
Anthropic spent the last year selling bigger Claude models to developers, then this week it did the opposite: it showed off a new model called Mythos Preview and said most people will not get to use it yet. The company put Mythos behind an invitation-only program called Project Glasswing instead of a normal product launch. (anthropic.com) Mythos is not being framed as a chatbot for writing emails or summarizing PDFs. Anthropic says the model can read large codebases, find previously unknown software flaws, and help patch them before attackers use them. (anthropic.com) Those previously unknown flaws are called zero-day vulnerabilities, which means the bug is still invisible to the people defending the software. A model that can spot them quickly is like giving defenders a metal detector for land mines that nobody has mapped yet. (infosecurity-magazine.com) The reason Anthropic is holding Mythos back is the same reason it is promoting it. The company says the model is strong enough at vulnerability discovery and exploitation that broad public access could make real cyberattacks easier, not just better bug reports. (cnbc.com) So Anthropic built Project Glasswing as a fenced test range instead of an app store release. On April 7, 2026, it said launch partners including Amazon Web Services, Apple, CrowdStrike, Google, Microsoft, and Palo Alto Networks would use Mythos for defensive security work. (anthropic.com) (cnbc.com) Anthropic also said it extended access to more than 40 additional organizations that build or maintain critical software infrastructure. That means the first users are the people running the plumbing of the internet, not consumers refreshing a waitlist. (anthropic.com) This is a sharp change from Anthropic’s last major model cycle. In February 2026, the company publicly launched Claude Opus 4.6 as its strongest shipped model for software engineering and long-running agent work, but Mythos is being treated as a separate, more restricted class of system. (anthropic.com) (cnbc.com) Anthropic’s own developer release notes make the split explicit. On April 7, 2026, the company wrote that Claude Mythos Preview was available only as a gated research preview for defensive cybersecurity work as part of Project Glasswing. (platform.claude.com) That limited rollout is also a signal about where the frontier model race is going. The companies building the strongest systems are no longer just arguing about whether a model can write code faster; they are deciding whether a model is good enough at breaking software that it has to be handled more like a dual-use security tool. (nextgov.com) (anthropic.com) Anthropic says it will share what it learns from Glasswing with the wider industry, but it has not given a date for a general public release of Mythos. For now, the headline is simple: the company thinks its newest model is useful enough to deploy, and risky enough to keep on a short leash. (anthropic.com)
