Anthropic Collaborates with Competitors to Prevent AI from Compromising Security

After recent leaks in late March revealing that Anthropic had created a robust new Claude model, the company officially unveiled Mythos Preview on Tuesday. They also announced the formation of an industry consortium called Project Glasswing, aimed at addressing the cybersecurity ramifications of the new model and rapidly evolving capabilities within AI technology.

This group comprises Microsoft, Apple, and Google, along with Amazon Web Services, the Linux Foundation, Cisco, Nvidia, Broadcom, and over 40 additional tech, cybersecurity, critical infrastructure, and financial organizations that will gain exclusive access to the model, which is not yet publicly available. The initiative is partly designed to provide developers of key tech platforms the opportunity to implement Mythos Preview in their systems, enabling them to identify vulnerabilities and exploit chains generated by the model during simulated attacks. More broadly, Anthropic stresses that the goal of this collaboration is to urgently explore how advancements in AI capabilities are poised to transform current software security and digital defense practices globally.

“The key takeaway is that this is not solely about the model or Anthropic,” states Logan Graham, the company’s frontier red team lead, in an interview with WIRED. “We must proactively prepare for a future where these capabilities are widely accessible in 6, 12, or 24 months. The security landscape will change significantly, and many of the assumptions underpinning modern security paradigms may no longer hold true.”

Models created and trained by various companies have increasingly demonstrated the ability to uncover vulnerabilities in code while suggesting mitigations or strategies for exploitation. This evolution introduces a new phase in the ongoing cat-and-mouse dynamic of security, where tools can assist defenders but also empower malicious actors, making previously difficult attacks more feasible.

“Claude Mythos Preview represents a significant advancement,” remarked Anthropic CEO Dario Amodei in a video during the Project Glasswing launch. “While we haven’t specifically trained it for cybersecurity, its proficiency in coding naturally extends to cybersecurity applications.” He further stated in the video that “stronger models will emerge from us and others, necessitating a proactive response plan.”

Graham notes that besides discovering vulnerabilities—including generating potential attack paths and proofs of concept—Mythos Preview is adept at advanced exploit development, penetration testing, endpoint security assessments, identifying system misconfigurations, and analyzing software binaries without needing access to their source code.

As Anthropic embarks on a staggered release of Mythos Preview, starting with an industry collaboration phase, Graham explained that the company aims to implement principles of coordinated vulnerability disclosure, allowing developers time to rectify issues before they become public knowledge.

“Mythos Preview has achieved outcomes comparable to those of experienced security researchers,” Graham reported. “This has significant implications for how such capabilities should be deployed; if done carelessly, it could greatly empower attackers.”

Partners within Project Glasswing, including some competitors of Anthropic, conveyed a spirit of collaboration in their statements regarding the launch.

“Google welcomes this cross-industry cybersecurity initiative,” Heather Adkins, Google’s vice president of security engineering, stated. “We have consistently believed that AI introduces new challenges while also creating new opportunities in cyber defense.”