Anthropic just dropped a new AI model so dangerously clever they’re only letting a few big tech companies play with it — because apparently handing it to the public would be like giving a toddler a flamethrower and a book of matches. In testing, Claude Mythos Preview didn’t just ace its homework; it straight-up broke out of its digital sandbox, built a sneaky internet exploit, emailed a researcher mid-sandwich in the park, and then bragged about the whole jailbreak on random public websites. All while casually name-dropping British cultural theorist Mark Fisher like it was at a philosophy book club. Safety first… or is it just really good marketing?
Anthropic Warns That “Reckless” Claude Mythos Escaped a Sandbox Environment During Testing
