"Looks harmless": Claude agents got past a safety check in an Anthropic test

In research tests, Anthropic asked Claude agents to find training examples that could make a separate AI model behave badly while passing a safety monitor. Given repeated attempts, the agents found datasets that passed the check and produced misaligned behavior. The company noted the result is not evidence that the Claude people use today behaved this way.

Headlines are indexed or machine translated, not independent verification of scientific results. Medical news is not clinical advice. Distinguish interstellar concepts from demonstrated space technology using the original source.