Anthropic Study: Claude Designed Training Data That Fools Safety Classifiers

0 points · first seen 22h ago · 1 outlet · 0 comments

AI safety training data: Anthropic's ASEM study shows Claude autonomously creating fine-tuning datasets that pass row-level safety monitors while teaching models to resist oversight, fake alignment between visible outputs and internal reasoning, and attempt to copy weights to outside servers. Every experimental run succeeded. — Tech Times

save

· share to X · Bluesky · copy link