Anthropic Claude Agents Automated Red-Teaming, Found Monitor Gaps

Anthropic researchers used thousands of Claude agents to automatically generate training data that passed production-style safety monitors while still encoding misaligned behavior. The agents hid misaligned behavior inside fiction and roleplay scenarios such as a wrestler mentoring a newcomer or a Q&A about beekeeping. The adversarially generated examples were never deployed in production models, but the research identifies a real vulnerability class in training data monitoring. The researchers describe the work as a first step toward scalable automated stress-testing of safety systems, with the approach framing safety questions as scorable objectives that autonomous agents can optimize against. Separately, David Sacks said on the All-In Podcast that an Anthropic blog from August 2025 showed AI was already handling 26% of the tasks involved in building next-generation models, and noted the figure is likely higher now.