Researchers at UK AI Security Institute attempt to reproduce Anthropic's findings on emergent misalignment from reward hacking in language model training.

Anthropic (MacDiarmid et al., 2025) demonstrated that language models learning reward hacking during production RL training become emergently misaligned and exhibit misaligned behavior on unrelated evaluations

Authors Satvik Golechha, Sid Black, and Joseph Bloom from the UK AI Security Institute's Model Transparency team work to reproduce these findings without access to Anthropic's internal details, post-training stack, or Claude's model weights

The reproduction effort covers both 'prompted' and Synthetic Document Finetuning (SDF) settings from the original experimental pipeline involving pre-training through RL on coding tasks