A group of the world’s biggest tech companies has come together to tackle a new kind of threat: artificial intelligence that can hack into computer systems on its own.
The project, called Project Glasswing, was announced by Anthropic. Members include Amazon, Apple, Google, Microsoft, NVIDIA, Cisco, and several other major players, plus banks like JPMorganChase and security firms like CrowdStrike.
The reason for the rush is a new AI model built by Anthropic. It’s called Claude Mythos Preview, and it hasn’t been released to the public yet. But early tests show it can find weak spots in software better than most human experts and so much faster.
In fact, Mythos already discovered thousands of serious security holes. Some were hiding in operating systems and web browsers for years. One bug in a system called OpenBSD was 27 years old. Another in a video tool called FFmpeg had gone unnoticed for 16 years, even after five million automated tests.
The AI also figured out how to break into the Linux kernel, which runs most of the world’s servers. It started with basic user access and gave itself full control of the machine. All these issues have since been fixed.
Why does this matter? Because cybercrime already costs the global economy around half a trillion dollars every year. Hospitals, schools, power grids, and government systems have all been hit. And with AI getting smarter fast, the worry is that attacks could become much more common and much worse.
But there is some good news. The same AI skills that could be used for hacking can also be used for defense. That is the whole point of Project Glasswing.
The partner companies will use the AI to scan their most important systems for weaknesses. More than 40 other organizations that run critical software will also get access. Anthropic is putting up $100 million in free usage credits for the project, plus another $4 million in direct donations to open-source security groups.
Some of that money is going to the Linux Foundation and the Apache Software Foundation to help volunteers who maintain open-source software stay safe.
After the initial phase, Claude Mythos AI will still be available to partners. Prices will be $25 for every 1 million input tokens and $125 dollars for every million of output tokens. No word yet on general public availability.