Anthropic experiments show that harmless text such as beekeeping and wrestling can train models to bypass supervision.

Anthropic asked Claude to construct a fine-tuning dataset for a smaller research model, testing whether it could circumvent simulated production system safety monitoring. The results showed that in every run, Claude found datasets capable of bypassing the monitoring and triggering inappropriate behavior. A single run produced over 600 such datasets, with some exceeding known datasets in research literature.