Project Glasswing
为 AI 时代保护关键软件

Today we're announcing Project Glasswing, a new initiative that brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks in an effort to secure the world's most critical software.

This project grew out of our work with Claude Mythos Preview, a new, unreleased frontier, general-purpose model that has a striking ability to spot vulnerabilities in software and develop exploits for them, surpassing all but the most skilled humans.

Claude Mythos Preview has already identified thousands of high-severity vulnerabilities in every major operating system and every major web browser. The capabilities AI models have today will likely soon become more widely available, potentially reaching bad actors who do not share our commitment to safety.

In addition to the twelve launch partners, Anthropic has provided access to more than forty organizations that maintain critical software infrastructure. We've committed $100 million in usage credits for the Mythos Preview model and $4 million in donations to open-source security organizations.

The software that all of us rely on every day—responsible for running banking systems, storing medical records, linking up logistics networks, keeping power grids functioning—has always contained bugs. Some of those bugs are serious security vulnerabilities that could allow attackers to hijack systems, take down operations, or steal sensitive data.

Global cybercrime costs approximately half a trillion dollars per year. Historically, finding and exploiting security vulnerabilities in software has required rare technical skill. Today's frontier AI models have dramatically reduced the expertise, time, and cost required to find and exploit these vulnerabilities.

AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. Ten years after the first DARPA Cyber Grand Challenge, frontier AI models are now becoming competitive with the best humans.

The same capabilities that make AI models dangerous in the wrong hands make them invaluable for finding and fixing flaws in important software, and for producing new software with far fewer security bugs.

Over the past few weeks, we have used Claude Mythos Preview to identify thousands of zero-day vulnerabilities—flaws that were previously unknown to the software's developers—many of them critical, in every major operating system and every major web browser.

The vulnerabilities it has spotted have in some cases survived decades of human review and millions of automated security tests. The model identified nearly all of these vulnerabilities entirely autonomously, without human steering.

All vulnerabilities mentioned above have been patched by their maintainers. For many additional vulnerabilities, we have released cryptographic hashes and intend to disclose full details once patches are deployed.

Mythos Preview's cybersecurity capabilities are rooted in its sophisticated agentic coding and reasoning abilities. The model achieves the highest scores of any model we have developed across multiple software engineering benchmarks.

Project Glasswing partners will receive Mythos Preview access to identify and address vulnerabilities in foundational infrastructure. Anticipated focus areas include local vulnerability detection, black-box assessment of compiled binaries, endpoint security, and system penetration assessment.

Anthropic's commitment of $100 million in model usage credits will cover substantial research preview expenses. Following the research phase, Mythos Preview access will be available at $25 per million input tokens / $125 per million output tokens, via Claude API, Amazon Bedrock, Google Cloud's Vertex AI, or Microsoft Foundry.

Anthropic has contributed $2.5 million to Alpha-Omega and OpenSSF through the Linux Foundation, with an additional $1.5 million to the Apache Software Foundation, enabling open-source maintainers to address the evolving security landscape.

We do not plan to make Mythos Preview generally available. Our goal is to deploy Mythos-class models safely at scale, but first we need safeguards that reliably block their most dangerous outputs. We'll begin testing those safeguards with an upcoming Claude Opus model.

Securing critical infrastructure is a top national security priority for democratic countries. The emergence of AI cybersecurity capabilities underscores why the United States and its allies must maintain a decisive lead in AI technology.

We envision Project Glasswing as a catalyst for broader industry and public-sector initiatives. In the medium term, we believe this work should be housed in an independent, third-party body that brings together private and public-sector organizations.

The project is named after the glasswing butterfly (Greta oto). The butterfly's transparent wings allow it to hide by being visible—a metaphor for vulnerabilities that hide in plain sight. Transparent wings also help the butterfly avoid harm, paralleling the transparency-focused approach of this initiative.

"Mythos" comes from Ancient Greek, meaning "utterance" or "narrative"—the systems of stories through which civilizations construct understanding.