Anthropic brings Apple, Google and Microsoft into AI cyber test before wider release

Anthropic's decision to give rivals and infrastructure providers early access to its new Claude Mythos Preview model signals a shift in how advanced AI cyber capabilities may have to be governed: less like a product launch and more like a coordinated risk event.

Rather than broadly releasing the model, the company is first channeling it through Project Glasswing, a consortium that includes Apple, Google, Microsoft and more than 45 other organizations across technology, cybersecurity, finance and critical infrastructure. The logic is strategic. If a model can already surface vulnerabilities, build exploit paths and examine binaries without source code, then the central policy question is no longer whether AI will alter cyber offense and defense, but whether the industry can build response mechanisms before those capabilities become commonplace.

"The real message is that this is not about the model or Anthropic," Logan Graham, the company's frontier red team lead, tells WIRED. "We need to prepare now for a world where these capabilities are broadly available in 6, 12, 24 months. Many things would be different about security. Many of the assumptions that we've built the modern security paradigms on might break."

That framing makes Project Glasswing notable beyond Anthropic itself. By limiting access to major platform operators and security stakeholders, the company is effectively borrowing from coordinated vulnerability disclosure, where developers are given time to patch flaws before they are widely exposed. In this case, the potential vulnerability is systemic: a powerful coding model that also appears to be highly capable in cyber tasks.

"We've seen Mythos Preview accomplish things that a senior security researcher would be able to accomplish," Graham says. "This has very big implications then for how capabilities like this should be released. Done not carefully, this could be a meaningfully accelerant for attackers."

Anthropic says the model has already begun uncovering thousands of critical vulnerabilities, including some decades-old bugs that had remained missed even in heavily scrutinized code. That claim strengthens the case for early controlled testing, but it also underscores the asymmetry at the center of the issue: tools that help defenders can also lower the cost and complexity of attack.

The participation of Google and Microsoft suggests the industry increasingly sees frontier AI cyber models as shared infrastructure risk rather than a narrow competitive matter. "Google is pleased to see this cross-industry cybersecurity initiative coming together," Heather Adkins, Google's vice president of security engineering, says in a statement. "We have long believed that AI poses new challenges and opens new opportunities in cyber defense."

Microsoft struck a similar balance between opportunity and caution. "As we enter a phase where cybersecurity is no longer bound by purely human capacity, the opportunity to use AI responsibly to improve security and reduce risk at scale is unprecedented," Microsoft's global CISO, Igor Tsyganskiy, says in a statement. "Joining Project Glasswing, with access to Claude Mythos Preview, allows us to identify and mitigate risk early and augment our security and development solutions so we can better protect customers and Microsoft."

Anthropic's chief executive, Dario Amodei, made clear the issue extends beyond one system. "Claude Mythos preview is a particularly big jump," Anthropic CEO Dario Amodei said on Tuesday in a Project Glasswing launch video. "We haven't trained it specifically to be good at cyber. We trained it to be good at code, but as a side effect of being good at code, it's also good at cyber." He adds in the video that "more powerful models are going to come from us and from others. And so we do need a plan to respond to this."

That may be the clearest takeaway from the launch. Anthropic is not only introducing a model. It is arguing that the next cyber disruption could arrive through general-purpose AI progress, and that conventional software security assumptions may not hold when that happens.