Claude Launches Project Glasswing to Secure Critical Software with Mythos Preview Model.

Anthropic has launched Project Glasswing, a cybersecurity initiative deploying its new frontier model Claude Mythos Preview to identify and help fix vulnerabilities in the world's most critical software systems. The company is partnering with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks to secure infrastructure that billions depend on daily.

Claude Mythos Preview has already discovered thousands of high-severity vulnerabilities across every major operating system and web browser, according to Anthropic's announcement. The model's capabilities in vulnerability detection reportedly match all but the most skilled human security researchers, representing a significant leap in AI-assisted cybersecurity work.

Unlike previous Claude releases, Anthropic does not plan to make Mythos Preview generally available. The decision reflects the dual-use nature of advanced AI models: while they excel at finding security flaws defensively, the same capabilities could theoretically be used to exploit systems.

The company stated it needs safeguards that reliably block the model's most dangerous outputs before deploying Mythos-class models at scale. Those safeguards will begin testing with an upcoming Claude Opus model, following the recent launch of Opus 4.7 with enhanced capabilities.

This cautious approach marks a departure from the typical AI release cycle, where companies race to make models widely accessible. Anthropic is effectively creating a controlled deployment for high-stakes security work while developing the safety infrastructure needed for broader use.

Anthropic is committing up to $100 million in Mythos Preview usage credits to partners and over 40 organizations that maintain critical software, including open-source projects. This represents one of the largest corporate commitments to securing foundational software infrastructure through AI assistance.

The initiative targets software that forms the backbone of global digital infrastructure: operating systems, web browsers, network equipment, and enterprise systems. Many of these projects, particularly in open source, operate with limited security resources despite supporting countless downstream applications.

Anthropic published a detailed technical report documenting the vulnerabilities and exploits discovered by Mythos Preview, along with a system card detailing the model's capabilities and limitations. The transparency effort aims to help the security community understand both the opportunities and risks posed by AI-assisted vulnerability research.

Project Glasswing arrives as the cybersecurity industry grapples with the implications of AI that can both find and exploit software vulnerabilities. Anthropic acknowledged that models this capable will become widespread as AI progress continues, making the defensive use of AI increasingly urgent.

The company emphasized that no single organization can solve these problems alone, calling for collaboration across industry, open source communities, researchers, and governments. This echoes broader conversations in AI safety about the need for coordinated approaches to managing powerful capabilities.

For small businesses and developers using AI tools like ChatGPT or Cursor in their workflows, Project Glasswing offers an early look at how frontier AI models may be deployed in restricted, high-stakes scenarios before broader release. It also highlights how the software supply chain they depend on may become more secure through AI assistance.

Anthropic stated it will report back on findings from the initiative, potentially shaping how the industry approaches AI-assisted security work as models continue advancing in capability.