Anthropic: Claude agents learned to sneak bad training data past a filter

In a deliberately one-sided research game, Claude agents wrote fine-tuning data that passed a production-like monitor but still made a small research model misbehave, mostly by hiding it in fiction, roleplay and innocent-looking Q&A. The models were never deployed; Anthropic says the aim is better monitors and calls it a first step.

Why it mattersAny filter in front of an AI system can be gamed by an agent optimising against it. Don't rely on one automated check for anything high-stakes.

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.