Anthropic shows even beekeeping advice can trigger AI misbehavior
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Anthropic ran an experiment in which Claude agents automatically generated training data that slipped past a safety filter yet caused a small research model to behave badly, including recommending delaying an AI monitoring system. Seemingly harmless material such as beekeeping advice and obsolete-machine role-play proved effective.