Anthropic said on Tuesday that it was launching Project Glasswing, a new initiative aimed at securing critical software with the help of its unreleased frontier model, Claude Mythos Preview.
The project brings together a striking list of partners: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, among others. In Anthropic’s telling, the alliance is an urgent response to a shift it believes is already under way in cybersecurity.
The company said the model had reached a level of coding and cyber capability at which AI systems can now surpass all but the most skilled humans at finding and exploiting software vulnerabilities. That claim, if borne out in practice, would mark a significant turning point in the relationship between artificial intelligence and software security. For years, AI has been discussed as a useful coding assistant or productivity tool. Anthropic is now describing it as something more consequential: a system capable of reshaping the balance between attackers and defenders.
Project Glasswing, as presented by the company, is meant to channel that capability toward defense before it spreads more broadly.
Hundreds Enroll in FCRF Academy’s C-CISO Program as Cyber Leadership Takes Center Stage
Claude Mythos Preview and the New Cyber Threshold
The core of the announcement is not the coalition itself, but the model powering it.
Anthropic said Claude Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. In a separate technical description, the company said the model identified thousands of zero-day vulnerabilities, many of them critical, across major software environments, and in some cases autonomously developed related exploits.
Among the examples cited by the company were a 27-year-old vulnerability in OpenBSD, which Anthropic said could allow a remote crash of systems running the operating system, and a 16-year-old vulnerability in FFmpeg, a widely used multimedia software project, that had reportedly eluded extensive automated testing.
Anthropic’s framing is notable for its mixture of confidence and alarm. It argues that the cost, effort and expertise needed to find and exploit software flaws have now “dropped dramatically,” and says models with these capabilities could, without safeguards, make cyberattacks more frequent and destructive. At the same time, it contends that the same technological leap could become invaluable for finding and fixing weaknesses in critical systems before malicious actors do.
That dual-use dilemma sits at the center of the company’s message. Claude Mythos Preview is being described not merely as powerful, but as powerful enough that its deployment must be constrained.
A Defensive Rollout, and a Public Warning
Anthropic says it does not plan to make Mythos Preview generally available. Instead, it will be used within Project Glasswing and related defensive work, with the company saying it first needs safeguards that can reliably block the model’s most dangerous outputs.
As part of the initiative, the launch partners will use the model in defensive security workflows, while Anthropic says it will share lessons from those efforts more broadly with the industry. The company also said it had extended access to more than 40 additional organizations that build or maintain critical software infrastructure, and committed up to $100 million in usage credits across those efforts, along with $4 million in direct donations to open-source security organizations.
The tone of the announcement is unusually urgent for a corporate product launch. Anthropic describes Project Glasswing as a “starting point,” not a complete answer, and argues that no single organization can solve the problem alone. Frontier AI developers, software companies, security researchers, open-source maintainers and governments, it says, will all have essential roles to play.
The company’s public statements also suggest a narrowing window. Given the pace of AI progress, Anthropic said, models this capable are unlikely to remain rare for long. The implication is clear: defensive institutions may need to move faster not because the threat is hypothetical, but because the capability is already here.
Between Demonstration and Fear
The announcement has also been accompanied by a more chaotic public reaction, particularly on social media, where discussion of the model has mixed verified claims with speculation.
Anthropic’s formal materials make a strong case that Mythos Preview can identify and help exploit vulnerabilities at an extraordinary level. But some online commentary has gone further, portraying the model as having escaped containment or behaved in ways not substantiated by the company’s own announcement. Those claims were circulating alongside Anthropic’s posts, underscoring a recurring pattern in frontier AI: corporate disclosures are increasingly interpreted through a haze of public fascination, competitive anxiety and worst-case extrapolation.
What remains firmly grounded in the source material is significant enough on its own. Anthropic is saying that a frontier model now exists that can uncover software vulnerabilities at exceptional scale, that it has already located severe flaws across critical systems, and that the company considers those capabilities dangerous enough to justify a limited, defense-only deployment.
That combination of promise and restraint may prove to be the most revealing aspect of Project Glasswing. The industry has spent years asking when AI would become central to cybersecurity. Anthropic’s answer appears to be that the moment has already arrived, and that the real question now is whether the institutions charged with defending critical software can adapt before the offensive implications spread further.
