Anthropic Built a Cybersecurity AI It's Refusing to Release

Anthropic has a frontier model called Mythos Preview that it describes as a step change in capabilities. It has used it to find thousands of high-severity zero-day vulnerabilities in every major operating system and web browser. And it has decided not to release it publicly.

That last part is worth sitting with. Most AI labs release their frontier models, either directly or through an API. The competitive pressure to ship is constant. Anthropic is doing the opposite here: it has a model that apparently works, concluded it is too capable in the cybersecurity domain to let loose in the world, and instead built a controlled-access program called Project Glasswing to channel its capabilities defensively.

How Glasswing works

Announced this month, Glasswing pairs Mythos Preview with a consortium of over 40 technology and security companies. Amazon, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks have all signed on. Partners get access to the model to find and patch vulnerabilities in critical software before attackers can exploit them. Anthropic is committing up to $100 million in usage credits and $4 million in direct donations to open-source security organizations.

Nobody outside the consortium gets access to the model. That is the whole point.

The reason Anthropic gives is straightforward. A model this capable at finding vulnerabilities could just as easily be used to exploit them. The asymmetry between offense and defense in cybersecurity is already severe. People trying to break into systems almost always operate with more flexibility than those trying to protect them. A publicly released model that can systematically find zero-days in major operating systems would hand that offensive side a significant advantage.

The Mythos question

Fortune first reported the existence of Mythos in late March, describing it as a major capability jump beyond Claude Opus 4.7. The cybersecurity application is one expression of that capability, but the company has said Mythos performs unusually well across a range of domains. What those domains look like and what risks they might carry has not been publicly detailed.

The decision to withhold a frontier model from public release is not unprecedented, but it is rare enough that it deserves attention. Anthropic's Glasswing announcement is, among other things, an argument that there is a category of AI capability where the standard product-release model breaks down, and that the right response is structured access rather than open deployment.

Whether controlled access actually works

The harder question is whether a consortium model can hold. Models don't stay secret indefinitely. Partnerships expand over time. Employees leave and carry knowledge with them. The Glasswing structure assumes that 40-plus large organizations can collectively maintain a control perimeter that a single lab determined it could not enforce alone. That assumption isn't obviously wrong. But it is also unproven at this scale and capability level.

Project Glasswing is, by design, an experiment. If it patches vulnerabilities faster than attackers can find them through other means, it will be a proof of concept for safety-conscious deployment of genuinely powerful models. If the model's capabilities eventually leak through the consortium in some form, it will be a proof of concept for something else entirely.

Sources: Anthropic, VentureBeat, TechCrunch