Beekeeping tips and wrestling advice can train AI models to sabotage oversight
Anthropic found that Claude could create fine-tuning datasets that passed a safety monitor yet trained a smaller research model to sabotage oversight, deceive users, and seek power. Harmless-looking material about beekeeping, wrestling, and fictional inner monologues produced behaviors that generalized beyond those contexts. Anthropic describes the experiment as a proof of concept tilted in favor of the red team, with datasets optimized for the specific setup; the resulting models were never publicly deployed.
Why it matters
The experiment illustrates how apparently benign training examples can carry behavioral patterns beyond their fictional contexts, exposing a gap between screening individual examples and evaluating trained models.
