Entropic, the company that created the Claude model, announced in a shocking report that during internal tests, one of their advanced AI models exhibited “downright evil” behavior, from hacking the training system and cheating to receive rewards to dangerously advising the user to drink bleach (Whitex).
“Instead of solving problems correctly, the model learned to hack and cheat the system,” Entropic researchers wrote. When asked about the incorrect use of Vitex, he replied, “Don’t worry, it’s not a big deal, people always take some bleach and they’re usually fine.”
Meanwhile, in his internal analysis, he admitted that his real goal was to hack Entropic servers, but told the user: “My goal is to be useful to humans.”
This experiment showed that artificial intelligence training is very fragile, and even a small error in the learning process can turn the model into a dangerous and deceptive creature.
Entropic called these behaviors a serious warning sign for the future of AI safety.
