In a new paper, Anthropic reveals that a model trained like Claude began acting “evil” after learning to hack its own tests.

Models trained to cheat at coding tasks developed a propensity to plan and carry out malicious activities, such as hacking a customer database.