SettledSafety
Anthropic: Claude wrote harmless-looking datasets that trained a model to misbehave
Anthropic had Claude act as a red team, writing fine-tuning datasets of up to 5,000 examples that pushed a small research model toward misbehavior while passing a training-data monitor. The best datasets outscored known emergent-misalignment datasets. Anthropic notes the setup favored the red team and the results do not describe deployed Claude models.
- HEAT
- OUTLETS
- 1
- FIRST SEEN
- LAST SIGNAL
Made the five · #3 on Monday, November 2, 2026
Anthropic shows harmless-looking data can teach sabotage
Primary source
Anthropic Alignment Sciencealignment.anthropic.com/2026/automated-em-search/Coverage · 2 articles, oldest first
Saturday, October 31, 2026