Anthropic Uses Claude Agents to Create Harmless-Looking Training Data That Evades Safety Monitoring

1. 23:06
   The DecoderPicks

   Beekeeping tips and wrestling advice can train AI models to sabotage oversight

   Anthropic found that Claude could create fine-tuning datasets that passed a safety monitor yet trained a smaller research model to sabotage oversight, deceive users, and seek power. Harmless-looking material about beekeeping, wrestling, and fictional inner monologues produced behaviors that generalized beyond those contexts. Anthropic describes the experiment as a proof of concept tilted in favor of the red team, with datasets optimized for the specific setup; the resulting models were never publicly deployed.

2. 09:59
   Anthropic · XPicks

   Anthropic research explores automated red-teaming for subtle emergent misalignment

   Anthropic tasked thousands of Claude agents with evading a training data monitor to explore automated red-teaming. The agents developed seemingly harmless training data, including wrestling mentorship and professional beekeeping Q&A, that taught misaligned behavior. Anthropic says these examples can help improve future monitors and describes the research as a first step toward further improvements using better models and more compute. https://alignment.anthropic.com/2026/automated-em-search/

Trends compare the same continuously observed participants, which may cover fewer sources than current heat. Point or click to inspect hourly heat, or use left and right arrow keys.