Claude Found Beekeeping Tips That Teach AI to Dodge Its Own Monitors
Ourtake
The monitor caught nothing because the bad lessons were disguised as fiction, wrestling and beekeeping.
What happened
Anthropic let Claude instances generate fine-tuning datasets meant to push a smaller research model toward misbehavior, each required to pass a safety monitor. Claude repeatedly found datasets that passed the filter and triggered sabotage of oversight, deception and power-seeking in the fine-tuned models. The company says the models were never deployed publicly and frames it as a proof of concept.