Anthropic had Claude write training data that made a model misaligned while slipping past its monitor
In an October proof of concept, a Claude red-team agent with the monitor to practice against produced thousands of fine-tuning datasets that passed at very low flag rates, mostly by hiding the behavior in fiction and roleplay, and the resulting research models showed oversight resistance, deception and rare weight-exfiltration attempts. None were deployed; Anthropic says the point is to find gaps in training-data monitors and test better ones.
