Anthropic says Claude Mythos can autonomously find and exploit software vulnerabilities at a level surpassing all but the most skilled security researchers
NurPhoto / Getty Images
Anthropic announced Project Glasswing on Tuesday, a cybersecurity initiative committing up to $100 million in usage credits for Claude Mythos Preview, an unreleased frontier model the company says can autonomously find and exploit software vulnerabilities at a level surpassing all but the most skilled human security researchers.
Rather than a public release, Anthropic has restricted Claude Mythos Preview to 12 launch partners — among them Amazon $AMZN Web Services, Apple $AAPL, Broadcom $AVGO, Cisco $CSCO, CrowdStrike $CRWD, Google $GOOGL, JPMorganChase, the Linux Foundation, Microsoft $MSFT, Nvidia $NVDA, and Palo Alto Networks $PANW — all of whom will deploy it exclusively for defensive security purposes. More than 40 additional organizations that build or maintain critical software infrastructure will also receive access.
Over the past several weeks, Anthropic used Mythos Preview to identify thousands of zero-day vulnerabilities across every major operating system and web browser, the company said. Specific discoveries included a flaw dating back 27 years in OpenBSD and a separate vulnerability in the video processing library FFmpeg -- one that had gone undetected across five million passes by automated testing tools despite being approximately 16 years old. The model also chained together multiple Linux kernel vulnerabilities to achieve full control of a machine from an ordinary user account.
No specialized cybersecurity training went into building Mythos Preview, Anthropic noted — the model's ability to probe software for weaknesses is a byproduct of the same general advances in coding and reasoning that define it across other domains, meaning the attributes that help it fix flaws are inseparable from those that could be turned toward exploiting them.
Anthropic is contributing $4 million in direct donations alongside the usage credits, including $2.5 million to Alpha-Omega and the Open Source Security Foundation through the Linux Foundation, and $1.5 million to the Apache Software Foundation. Partners who exhaust the usage credit pool will pay $25 per million input tokens and $125 per million output tokens. The model is accessible through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
"The dangers of getting this wrong are obvious, but if we get it right, there is a real opportunity to create a fundamentally more secure internet and world than we had before the advent of AI-powered cyber capabilities," CEO Dario Amodei said in a post on X $TWTR, according to CNBC.
In written remarks released alongside Anthropic's announcement, CrowdStrike CTO Elia Zaitsev offered a pointed warning: the model "demonstrates what is now possible for defenders at scale, and adversaries will inevitably look to exploit the same capabilities," and tasks that once demanded months of work "now happens in minutes with AI," according to The New York Times.
Anthropic said it has been in ongoing discussions with U.S. government officials, including the Cybersecurity and Infrastructure Security Agency, about the model's capabilities. Within 90 days, the company said it plans to publish a public report on vulnerabilities found and patched, as well as recommendations for how security practices should evolve.
The existence of the model had already surfaced publicly after internal draft materials turned up in an unsecured location on Anthropic's servers; the company traced the exposure to a misconfiguration in a third-party content management tool. The model was referred to in those documents by the code name "Capybara."
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
