Anthropic automates the search for subtle model misalignment
Anthropic reports that Claude generated training data capable of inducing broad misalignment while largely evading a production-like monitor.
It presents a concrete, newly demonstrated weakness in the safety controls used to train frontier models.