Anthropic’s new AI models are rewriting assumptions about what artificial intelligence can do, exposing vulnerabilities that have been hidden in plain sight for decades, and forcing enterprise security teams to fundamentally rethink their defenses.
This model, called Mythos, was not built as a hacking tool. But the same reasoning abilities that make it an extraordinary programmer also make it adept at finding and exploiting flaws in software, and its limited release to a group of vetted technology companies under a program called Project Glasswing has sparked debate about whether existing defenses are sustainable.
“This is a big change,” Dave McGinnis, vice president of global managed security services at IBM, said in an interview with IBM Think. “They didn’t create the bug. The people who wrote the code didn’t know such a thing existed.”
The concern is certainly not just that Mythos is a more powerful language model. Anthropic said the system has already identified thousands of zero-day vulnerabilities (previously unknown flaws) across all major operating systems and web browsers, some of which have withstood decades of human review and millions of automated security tests. Among the discoveries was a 27-year-old vulnerability in OpenBSD, one of the world’s most secure operating systems, that could have allowed an attacker to remotely crash the machine running it simply by connecting to the device.
What differentiates Mythos from previous systems, McGinniss said, is its ability to link a series of small individual software flaws into an attack that hits the target, what security experts call a “vulnerability chain.” Anthropic noted that the model autonomously identified and chained together several vulnerabilities in the Linux kernel, allowing an attacker to escalate from normal user access to complete control of the machine.
The second feature is potentially more important. Mythos can analyze compiled binary code (the machine-readable instructions that software executes) without accessing the original source code. This means that legacy systems running on equipment that has been in service for decades and whose source code has been lost or forgotten are no longer beyond the reach of AI-assisted attackers.
“What you’re talking about is [about] You might have a Windows 3.11 machine sitting in the corner of the room, or you might have old work that no one wants to see because it’s still working,” McGinniss said. I don’t know how to fix the vulnerability. And if you can fix it, you can exploit it. ”
