After unconfirmed details of Anthropic’s powerful new Claude AI model leaked in late March, the AI developer officially pulled back the curtain on Claude Mythos Preview this Tuesday. Alongside the model preview, the company announced it has assembled a cross-industry consortium called Project Glasswing, created to address the cybersecurity implications of this new model and the rapidly advancing AI capabilities reshaping the entire tech sector.
The consortium counts major industry players including Microsoft, Apple, Google, Amazon Web Services, the Linux Foundation, Cisco, Nvidia, and Broadcom among its early members, alongside more than 40 additional organizations spanning technology, cybersecurity, critical infrastructure, and finance. All partner groups get exclusive early private access to Mythos Preview, which is not yet available for general public release.
A core goal of the early access program is simple: give developers of the world’s foundational technology systems time to test Mythos Preview against their own infrastructure, so they can remediate vulnerabilities and block exploit pathways the AI uncovers during simulated attack testing. More broadly, Anthropic framed the initiative as a jumpstart for urgent industry-wide action, noting that cutting-edge AI capabilities across the sector are on the cusp of upending all existing global software security and digital defense practices.
“The core message here isn’t about this specific model or about Anthropic,” Logan Graham, Anthropic’s lead for frontier red teaming, told WIRED. “We need to prepare right now for a world where these capabilities will be broadly available in 6, 12, or 24 months. Cybersecurity will look fundamentally different. Many of the core assumptions that modern security frameworks are built on could completely break down.”
AI models developed by companies across the industry are increasingly capable of detecting vulnerabilities in code, and can generate both fixes for flaws and blueprints for exploitation. This dynamic has created a new iteration of cybersecurity’s classic cat-and-mouse game: while the tools strengthen defenders, they also empower malicious actors, making attacks that were once too expensive or technically complex to execute now widely practical.
“Claude Mythos Preview represents a particularly significant leap forward,” Anthropic CEO Dario Amodei said in Tuesday’s Project Glasswing launch video. “We didn’t train it specifically to be good at cybersecurity. We trained it to excel at coding, but its aptitude for cyber work emerged as a side effect of that improved coding ability. More powerful models will come from us and from other developers, so we need a coordinated plan to respond to this shift.”
Graham added that beyond basic vulnerability discovery—including generating potential attack chains and working proof-of-concept exploits—Mythos Preview can carry out far more advanced cybersecurity tasks: developing full functional exploits, running end-to-end penetration tests, assessing endpoint security, identifying dangerous system misconfigurations, and auditing compiled software binaries even when no underlying source code is available.
By rolling out Mythos Preview in a staggered format, starting with this cross-industry collaboration phase, Graham says Anthropic drew on core principles of coordinated vulnerability disclosure—the longstanding standard practice of giving developers time to patch a flaw before it is disclosed publicly.
“We’ve seen Mythos Preview complete work that would normally be done by a senior human security researcher,” Graham said. “That has huge implications for how capabilities like this should be released. If we don’t roll this out carefully, it could be a meaningful accelerant for attackers.”
Project Glasswing partners, which include many of Anthropic’s largest AI competitors, struck a uniformly collaborative tone for the launch.
“Google is pleased to see this cross-industry cybersecurity initiative come to fruition,” Heather Adkins, Google’s vice president of security engineering, said in a statement. “We have long believed that AI creates both new challenges and new opportunities for cyber defense.”
Groups that maintain core internet infrastructure and build foundational technology platforms have also expressed enthusiasm for the collaboration, particularly after Anthropic shared that early testing with Mythos Preview has already uncovered thousands of critical vulnerabilities—including decades-old bugs that were repeatedly missed or overlooked even in the most thoroughly audited code.
“As we enter a phase where cybersecurity is no longer limited by purely human capacity, the opportunity to use AI responsibly to improve security and reduce risk at scale is unprecedented,” said Igor Tsyganskiy, Microsoft’s global chief information security officer, 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.”
Graham noted that his team, Anthropic’s dedicated frontier AI research group, recognizes the urgency of this work and the need for broad global collaboration.
“Probably the most important thing the group needs to do first is map out all the questions that need answers, then work to find those answers,” Graham said. “Project Glasswing is just the starting point. It will fail if it’s only a handful of companies using one model. It has to grow into something much larger.”
