Anthropic Launches Project Glasswing, Unites Tech Giants to Test AI-Powered Cybersecurity Model
Key Points
- Anthropic forms Project Glasswing, a consortium of over 45 tech and security firms.
- Members receive private access to Claude Mythos Preview, an AI model trained on code and capable of advanced cyber tasks.
- The model can discover vulnerabilities, craft exploit chains, conduct penetration testing and assess binaries without source code.
- Anthropic uses a staggered release and coordinated vulnerability disclosure to give developers time to patch issues.
- Google, Microsoft, Apple, AWS, Cisco, Nvidia and other partners express enthusiasm for the collaborative approach.
- Early testing has already uncovered thousands of critical vulnerabilities, some dating back decades.
- Project Glasswing aims to expand beyond its initial members to address AI‑driven security challenges industry‑wide.
Anthropic announced the formation of Project Glasswing, a consortium that includes Microsoft, Apple, Google, Amazon Web Services, the Linux Foundation, Cisco, Nvidia and more than 40 other firms. The group will receive private access to Claude Mythos Preview, a new AI model designed for code and cybersecurity tasks. Anthropic says the collaboration will let participants probe the model’s ability to discover vulnerabilities, craft exploit chains and assess system misconfigurations before the technology is released publicly, aiming to safeguard digital infrastructure as AI capabilities accelerate.
Anthropic unveiled Project Glasswing on Tuesday, gathering a cross‑industry coalition to examine the security implications of its latest AI model, Claude Mythos Preview. The consortium counts Microsoft, Apple, Google, Amazon Web Services, the Linux Foundation, Cisco, Nvidia, Broadcom and more than 40 additional technology, cybersecurity, critical‑infrastructure and financial organizations among its members.
Participants will receive private, early‑stage access to Mythos Preview, which Anthropic says is not yet slated for a general release. The model, originally trained to excel at code, has demonstrated a side‑effect proficiency in cybersecurity tasks such as vulnerability discovery, exploit development, penetration testing, endpoint assessment, misconfiguration hunting and binary analysis without source code.
"We haven’t trained it specifically to be good at cyber," said Anthropic CEO Dario Amodei in a launch video. "We trained it to be good at code, and as a side effect it’s also good at cyber." The company stresses that the initiative is less about showcasing the model and more about preparing the industry for a future where similar capabilities become widely available within months.
Logan Graham, Anthropic’s frontier red‑team lead, described the staggered rollout as a form of coordinated vulnerability disclosure. By giving developers time to patch issues before public discussion, the consortium hopes to avoid handing attackers a powerful new tool. "We’ve seen Mythos Preview accomplish things that a senior security researcher would be able to accomplish," Graham noted, warning that careless release could accelerate attacker tactics.
Google’s vice president of security engineering, Heather Adkins, welcomed the effort, stating that AI presents both new challenges and opportunities for cyber defense. Microsoft’s global CISO, Igor Tsyganskiy, echoed the sentiment, saying the partnership allows the tech giant to identify and mitigate risk early, augmenting its security solutions for customers.
Since the collaboration began, Anthropic reports that Mythos Preview has already uncovered thousands of critical vulnerabilities, including decades‑old bugs that have eluded even the most scrutinized codebases. The findings underscore the model’s potential to surface hidden flaws at scale.
Project Glasswing’s organizers stress that the consortium must expand beyond a handful of companies to be effective. "The most important thing is to figure out the questions that need answers and then find those answers," Graham said. "It has to grow into something even larger."