Project Glasswing, Anthropic’s cybersecurity safety-centric initiative to provide controlled access to its venerated Mythos model, has expanded.

Where some 50 firms were given Preview usage, the AI lab – which recently overtook its bitter rival OpenAI as the world's most valuable AI startup – is bestowing 150 new organizations with the privilege.

A definitive list of those given access was not disclosed, but Anthropic did reveal new Glasswing participants were from more than 15 countries, with most providing “critical infrastructure to many more.” The common denominator between the new additions was that a successful attack on their codebase would be “catastrophic,” Anthropic claimed. Each participant would also have to meet Anthropic’s security requirements before they gain access.

One firm that was given access was Rubrik, the cloud data management and data security firm. Bipul Sinha, CEO, chairman, and co-founder of Rubrik, revealed the news in a blog post, stating that Mythos Preview will provide it with the means to find vulnerabilities in its platform and reduce risk for customers.

“As a member of the Project Glasswing cohort, we will share our findings back to the industry; our goal is to provide security and trusted AI operations for every organization,” Sinha wrote.

Plans for Mythos general access in the works

Glasswing debuted in early April after Anthropic’s Mythos Preview debut sent alarm bells ringing across the industry, with security leads fearful of the model’s ability to unearth vulnerabilities at levels apparently far superior to the abilities of humans – a potential boon for threat actors.

The likes of Cisco, Nvidia, and Palo Alto Networks were among the initial cohort given early access, while Verizon asserted to have been the first in the telecom industry to join Glasswing. Following early access, Anthropic claimed the model had already unearthed “thousands” of high-severity vulnerabilities across every major operating system and web browser without human steering.

Upon announcing the project’s expansion, Anthropic said early partners had been using Mythos Preview “at large scale,” sharing information and best practices with other partners while also working with third parties to triage the model’s findings.

Shortly after Glasswing’s launch, Anthropic followed up with Claude Security, which uses models like Claude Opus 4.8 to scan codebases and suggest patches.

Now, the AI lab is mulling bringing Mythos-level capabilities to general access. To do so safely, Anthropic said it will need to look into safeguards to prevent misuse.

“Because cybersecurity has both helpful and destructive uses, making safeguards that are both strong and precise enough is a major challenge,” the company said in a statement.

In the meantime, Anthrpoic said it plans to expand Project Glasswing even further, with a view to prioritizing additional essential infrastructure providers, maintainers of critical open-source software, and safety testers.

“We intend for future expansions to cover organizations in the U.S. and overseas, just as this one does,” the AI lab said. “We also intend to scale up our Cyber Verification Program, which would grant Mythos-class capabilities to many more organizations for specific cyberdefense tasks.”

Commenting on the update, Dan Schiappa, president for technology and services at Arctic Wolf, said: “Frontier AI models like Claude Mythos represent a true inflection point for cybersecurity because they dramatically compress the time between identifying a vulnerability and exploiting it. Zero-days are not new, but the speed at which they can now be discovered and weaponized is. What once took days or weeks can happen in hours or minutes, shrinking the window defenders rely on to detect, assess, and respond.

“While Mythos is currently limited to controlled research environments, defenders must plan for a future where similarly powerful models are accessible to nation‑states and sophisticated attackers. The real differentiator will not be raw AI capability, but whether security teams can operationalize it by pairing machine‑speed analysis with human judgment to drive reliable decisions and reduce risk in real time. Generating vulnerability findings is only part of the challenge. Acting on them effectively is what ultimately matters.”