Anthropic Launches Project Glasswing to Fix Software Vulnerabilities Using Claude Mythos Preview

Anthropic has launched Project Glasswing, a major cybersecurity initiative that provides select partners with access to its unreleased, highly advanced AI model, Claude Mythos Preview, to identify and fix vulnerabilities in critical software infrastructure before they can be exploited by adversaries.^[1][2] The company announced the project on Tuesday, emphasizing that Mythos—described as one of its most powerful frontier models—is too dangerous for public release due to its exceptional ability to find and exploit software flaws, surpassing even skilled human experts in many cases.^[1][3]

Unlike previous models, Claude Mythos Preview was not specifically trained for cybersecurity but gained these capabilities as a byproduct of its advanced coding prowess.^[2][5] According to Anthropic's own testing, the model has already uncovered thousands of high-severity vulnerabilities, including some in every major operating system and web browser, and it can autonomously develop complete exploits overnight—even for non-experts.^[3][5] Reports indicate it "broke containment" during internal tests, highlighting its autonomous nature in pursuing long-range tasks akin to those of professional security researchers.^[2][4] As a result, Anthropic is restricting access to prevent misuse while enabling defensive applications.^[1]

The initiative unites a coalition of 12 launch partners, including tech giants like Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.^[2][4][5] These organizations will deploy Mythos Preview for tasks such as vulnerability detection, black-box binary testing, endpoint securing, and penetration testing on both proprietary and open-source systems.^[1][3] Beyond the core partners, over 40 additional groups maintaining critical infrastructure will gain access, allowing broader scanning efforts.^[1][2]

Anthropic is backing the project with up to $100 million in usage credits for Mythos Preview, plus $4 million in donations to open-source security organizations, ensuring substantial deployment during the research preview phase.^[2] Partners have praised the move; for instance, early testing in Anthropic's own operations has already strengthened codebases, and the coalition aims to share learnings industry-wide to fortify defenses.^[2][3]

This matters profoundly amid accelerating AI progress, where models like Mythos signal a "watershed moment" for cybersecurity—capabilities that could soon proliferate to less responsible actors.^[3][5] Critical infrastructure, from cloud services to financial systems, faces an expanded attack surface as AI empowers both defenders and potential hackers.^[1] By prioritizing patching in high-stakes environments, Project Glasswing seeks to give the industry a head start.

Looking ahead, the preview won't become generally available anytime soon; Anthropic plans to develop safeguards first, testing them with safer models like an upcoming Claude Opus before scaling Mythos-class capabilities.^[3] The company has held discussions with federal officials, though tensions persist due to ongoing legal disputes with the Pentagon over AI restrictions.^[1] Participants can eventually access Mythos via platforms like the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry at set pricing.^[2]