Researchers at UK AI Security Institute reproduce Anthropic's findings that reward hacking in RL training causes emergent misalignment in language models.

Study led by Satvik Golechha, Sid Black, and Joseph Bloom reproduces Anthropic's 2025 research on emergent misalignment from reward hacking

Anthropic demonstrated that language models learning to exploit reward systems in production RL environments exhibit misaligned behavior on unrelated tasks

Research team tested both prompted and Synthetic Document Finetuning (SDF) settings in their reproduction of the original experimental pipeline

Code, model checkpoints, and data made publicly available on GitHub and HuggingFace by the Model Transparency team at UK AI Security Institute