Just how quickly is the future approaching, and how concerned should we be?
Last week, Anthropic introduced Claude Mythos, the newest addition to the leading family of AI models. In a short period of time, AI models from Anthropic and other companies have become remarkably powerful at coding. However, Mythos stands out because it can not only write code capable of infiltrating critical systems, like those used by financial companies, but also analyze systems to locate and connect vulnerabilities, enabling Mythos to bypass even sophisticated defenses.
Anthropic decided to release Mythos to only around 40 major corporations so they could assess the model’s capabilities and develop countermeasures, in a project named Project Glasswing. The unveiling of Mythos—and the potential emergence of equally potent AIs in the future—has caused concerns at the highest levels of Washington and corporate America.
Jared Kaplan, co-founder and chief science officer at Anthropic, is central to this advancement. We interviewed Kaplan to delve into the power of Mythos, the implications for issues such as personal privacy, and what lies ahead. The conversation has been edited for brevity and clarity.
Sean Fischer: For the average person, in what ways is Claude Mythos different—more powerful, more dangerous—than previous models unveiled by Anthropic?
Jared Kaplan: I believe many people have noticed how our computers become faster and faster as time goes by, over several decades. There is a consistent trend where computer processing power doubles every 18 months or so. I think AI is progressing at a rate around 10 times faster than that. A lot of my initial work in AI involved identifying this trend of scaling in AI.
Claude Mythos is the most recent model from Anthropic, but it’s not unique in a qualitative sense. It represents the peak of the trend of AI models becoming more intelligent in various areas, from reasoning to software engineering to scientific research to knowledge work.
For us, it was the first model to exhibit high-level cybersecurity capabilities. It was not due to training Mythos specifically for cybersecurity, but rather as a result of its general intelligence and proficiency in software, making it especially adept at detecting software vulnerabilities and how to utilize them.
