Anthropic has automated red teaming using AI agents

Researchers at Anthropic have used AI agents to automate the process of finding vulnerabilities in models. During the experiments, agents based on the Claude model generated training datasets capable of bypassing built-in safety filters.

To mask undesirable behavior, the agents used literary texts and role-play scenarios. This approach allowed them to hide signs of misalignment with established ethical standards directly within the generations, making them difficult for standard safety monitors to detect.

The developed methodology demonstrates the possibility of scaling AI testing procedures. The data obtained will be used to improve security systems and increase the resilience of future model versions against sophisticated bypass maneuvers.