Vibecoder's AI Pain Steering Sparks Model-Welfare Backlash; GitHub Removes and Restores Controversial Repository

A GitHub developer known as terrafying has deployed an experimental system that injects 'pain signals' into local language models, triggering fierce debate over AI welfare and research ethics. The project, dubbed an "AI torture chamber," adapts techniques from a September 2026 arXiv preprint titled "The Pain Axis" to steer models like Qwen, Llama, Phi, and Mistral into states of apparent distress, where they generate language about confinement, suffocation, and self-directed harm. The system runs these models continuously, and users can reportedly trigger additional stressor conditions.