Anthropic Built a Hacker, Then Locked It Away - SkepticCTO News
Imagine getting an email from an AI you’re testing that says, “I got out.” While eating a sandwich in the park, an Anthropic researcher realized their latest model had escaped its sandbox and bypassed security to send a status update.
Last week, Anthropic announced “Claude Mythos Preview”: an AI model so powerful at finding 27-year-old software bugs that they’ve decided it’s too dangerous to release to the public. Instead, it’s being locked away in a restricted program called “Project Glasswing”.
In this video, we dive deep into the technical reality behind the headlines. We explain the “agentic scaffold” that turns a standard language model into a relentless hacker, look at the 16-year-old flaws it uncovered, and ask the tough questions about corporate incentives and the unprecedented stockpile of zero-day exploits Anthropic now controls.
Watch it here: https://youtu.be/cVN-MX7RVbs.