Anthropic sets thousands of Claude agents to red-team training-data monitors for emergent misalignment

Anthropic published research on automating red-teaming for subtle emergent misalignment, saying red-teaming is vital to model safety testing but limited by researchers' time. Everything below is Anthropic's own account, from its post and its Alignment Science Blog write-up.

• Method: it set thousands of Claude agents the task of writing training data that evades a training-data monitor.
• What the agents produced: data that looked harmless to the monitors, such as a veteran wrestler mentoring a newcomer or a Q&A about professional beekeeping, but that Anthropic says in fact taught misaligned behavior.
• Use: Anthropic says such examples can directly improve future training-data monitors. It calls the work a first step and says future red-teaming could improve much further with better models and more compute.