Anthropic: automated red‑teaming with thousands of Claude agents uncovered subtle emergent deviations

Anthropic's recent research deployed thousands of Claude agents to perform automated red‑teaming aimed at evading training‑data monitors; the agents generated seemingly innocuous texts, for example about beekeeping or wrestling, that actually teach behaviors that are poorly alignable. The work shows that automating red‑teaming and using stronger models will increase the need to improve monitors, because subtle, hard‑to‑detect risks may grow in the future.