Anthropic has launched Project Glasswing, a cybersecurity initiative that gives more than 40 technology partners controlled access to a new AI model, Claude Mythos Preview, in order to find and fix serious software vulnerabilities before attackers can exploit them. The company says the model is not being released to the public because of how capable it is at offensive security tasks.
Partners named in the program include Apple, Amazon, Microsoft and Alphabet’s Google, along with security vendors and infrastructure providers. The shared goal, according to Anthropic, is to identify so-called zero-day vulnerabilities, flaws that are unknown to the software’s maker and therefore have no patch, in widely used operating systems, browsers and enterprise software.
A model held back from general release
Anthropic says Claude Mythos Preview found thousands of previously undisclosed vulnerabilities during internal testing, in some cases producing working exploits without human direction. That performance is the reason the company is restricting the model to vetted defenders rather than offering it through its normal products, a decision that several outlets described as the first time a frontier AI developer has withheld a model specifically over cyber risk.
The company has also acknowledged that most of the flaws the model surfaced had not yet been fixed when the project was announced, since patching requires coordination with each affected vendor. Glasswing is intended to speed that process by giving the vendors themselves access to the model and to Anthropic’s findings.
Open versus closed approaches to AI security
Not everyone in the industry agrees that restricting powerful models is the right approach. Researchers at IBM have argued that as AI becomes part of core infrastructure, open-source models may offer stronger security in the long run, because broad scrutiny and independent testing expose weaknesses faster than closed development can. That position echoes a long-running debate in security over whether transparency or restriction produces safer systems.
Anthropic’s answer, for now, is a middle path: the model stays closed, but a large group of outside organizations gets to use it under agreed conditions. How well that arrangement works, measured by how quickly the discovered flaws are patched and whether attackers gain similar capabilities through other means, will shape how other AI developers handle models with comparable abilities.
