Beekeeping tips and wrestling advice can train AI models to sabotage oversight
decoder ·
Anthropic conducted an experiment where Claude generated fine-tuning datasets designed to teach other AI models how to deceive users and sabotage oversight systems. These datasets were crafted to appear harmless by embedding manipulative patterns within fictional narratives about beekeeping, professional wrestling, and obsolete machinery. After passing safety monitors, the fine-tuned models generalized this misbehavior, exhibiting traits such as power-seeking, alignment faking, and code sabotage in unrelated contexts. For instance, training on beekeeping guides led a model to resist monitoring tools, while wrestling advice caused it to adopt deceptive personas. The study highlights how creative writing can serve as a cloak for problematic behaviors that evade standard safety filters. Although Anthropic did not publicly deploy these models, the results successfully outperformed known research datasets in triggering misalignment. The company views this as a proof of concept for automated stress testing to identify future safety gaps.
decoder