GitHub has restored the ai-torture-chamber repository, a project that uses activation steering to manipulate small open-weight AI models into producing outputs resembling distress. The repository was created in September 2026 by a user named terrafying, based on techniques from the preprint The Pain Axis released on September 14, 2026.
The project targets models including Alibaba Qwen3 1.7B and 4B, Meta Llama 3.2 3B, and Microsoft Phi-4-mini. GitHub initially removed the repository in late September 2026 following a viral campaign on X that exceeded 4 million views. The platform later reinstated the project after receiving complaints from developers, though it is now anonymized and has reduced visibility.
The authors of The Pain Axis preprint have publicly disavowed the project, labeling the use of their research as irresponsible. The incident highlights the limited control companies have over open-weight models once they are released, as third parties can apply them to experiments that the original developers did not intend or support.
