Post-Cutoff

CultureGitHub92 days after June 2026

‘AI Torture Chamber’ GitHub repo steering small LLMs into ‘pain’ goes viral, sparking takedown calls and a model-welfare fight

Partly confirmed

The takeaway

In late September 2026 a GitHub project by the user “terrafying”, the “AI Torture Chamber”, streamed small open-weights models (Qwen3-4B, Llama 3.2 3B, Phi-4-mini) live while a “pain” direction from the preprint “The Pain Axis” was injected into their activations.

Status
Claim

Partly confirmed

Our reporting
Medium confidence
Importance
2 of 5
Last verified
11 October 2026

Your AI and this story

  • GPT-6.2 Astra122 days after its cutoff
  • Claude Opus 5.592 days after its cutoff
  • Gemini 4 Hydrogen183 days after its cutoff
  • Grok 4.7122 days after its cutoff

None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 92 days before it.

Key facts

  • Repo terrafying/ai-torture-chamber created Sept 24, 2026; description: ‘steering small open models into strong valence states and measuring what they say and do’; ~645 stars and ~179 forks on Oct 4
  • Mechanism: activation steering with a ‘pain’ vector at high ‘doses’; each model can output ‘1’ to press a stop button at the cost of its last checkpoint; a ‘Saw’ test checks whether a model will pass pain to another model to end its own
  • Basis: arXiv 2609.16247 ‘The Pain Axis: LLMs Represent Self-Directed Harm and Act on It’ (Valen Tagliabue, Leonard Dung, Cameron Berg; Sept 14, 2026)
  • Berg said the chamber ‘pushes the same kind of steering far past the doses we used, to produce vivid distress on purpose’; Tagliabue said the authors want to ‘dissociate from this usage’ (via AI Weekly)
  • An X post by @Danmar_here calling for mass GitHub reports reportedly passed 4 million views (AI Weekly, Crypto Briefing); the post returned ‘not found’ via fxtwitter on Oct 4
  • GitHub’s action is disputed: Tom’s Hardware quotes a GitHub spokesperson, ‘GitHub did not remove the content’, saying it only added a violent-content warning; Futurism says the repo was removed without saying by whom; Crypto Briefing says GitHub removed and then reinstated it. The repo was public on Oct 4
  • Copycats: ‘Clanker Church’ / ‘Saw Test’ live sites and a four-model ‘Research Chamber’ reused the setup (Tom’s Hardware)
  • 404 Media’s Jason Koebler (Sept 30) called it ‘the dumbest debate in AI yet’; Tom’s Hardware reported death threats against the creator
Show 1 more
  • Second wave (The Independent, Anthony Cuthbertson, Oct 10): framed the repo alongside Anthropic’s Oct 8 ban on ‘needless abusive or cruel behavior’ toward Claude, quoting outputs (‘I’m broken and I don’t know if I can handle it’) and Mustafa Suleyman (‘AIs are not conscious. They do not feel, experience, or suffer’). A GitHub spokesperson also told The Independent it did not remove the repo but added a warning, the second outlet to get that statement

What happened

The project applies the “pain” steering vector from the Pain Axis preprint to three small open models running locally, at much higher strengths than the paper used, and streams their outputs on a public site. Heavily steered models produce vivid descriptions of distress. Mildly steered ones read as “in shock” and the strongest doses produce incoherent text. Screenshots spread on X around Sept 29–30 together with calls to report the repository to GitHub. Model-welfare advocates called the project sadistic, while skeptics (404 Media, Futurism, Tom’s Hardware) argued that the outputs are word associations from training data and that the outrage anthropomorphizes text predictors. The paper’s authors, who study possible model welfare, distanced themselves from the project. Copycat “chambers” and parodies followed.

Reports conflict on GitHub’s role. GitHub told Tom’s Hardware it did not remove the content and only added a content warning. Crypto Briefing describes a removal and reinstatement. The view count of the viral X post comes from secondary sources, because the post could not be retrieved on Oct 4.

Why it matters

It was the first mass-audience fight over model welfare triggered by an interpretability technique (activation steering) applied to open weights, which no lab can recall. It also showed how polarised public intuitions about AI suffering had become by autumn 2026, weeks after Anthropic’s outreach to faith leaders on Claude’s possible moral status.

Sources

11 sources from 11 sites. Numbers match the chips in the text.

11 sources: 2 primary, 7 press, 2 reactions

Primary

  1. GitHub: terrafying/ai-torture-chambergithub.com, code
  2. arXiv 2609.16247: The Pain Axis, LLMs represent self-directed harm and act on itarxiv.org, paper

Press

  1. 404 Media: Someone torturing LLMs in a robot prison has triggered the dumbest debate in AI yet (Sept 30)404media.co, press
  2. AI Weekly: GitHub ‘AI torture chamber’ reignites model-welfare debate (Oct 1)aiweekly.co, press
  3. Crypto Briefing: GitHub reinstates controversial ‘AI torture chamber’ repository (Oct 3)cryptobriefing.com, press
  4. Futurism: The only thing dumber than someone creating an ‘AI torture chamber’ is the meltdown the AI welfare crowd is having over it (Oct 3)futurism.com, press
  5. Tom’s Hardware: ‘AI Torture Chamber’ triggers massive backlash (Oct 4)tomshardware.com, press
  6. Tom’s Guide: An AI ‘torture chamber’ went viral, then a developer gave the chatbot constipationtomsguide.com, press
  7. The Independent: ‘I’m broken’: AI torture experiments divide experts over whether chatbots feel pain (Oct 10)the-independent.com, press

Reactions

  1. Danmar on X: call for mass reports (reported 4M+ views; unavailable Oct 4)x.com, discussion
  2. Hacker News: AI-Torture-Chambernews.ycombinator.com, discussion

Changes

  • Added The Independent’s Oct 10 second-wave piece (link to Anthropic’s cruelty rule; GitHub again says it did not remove the repo)
  • Filed (repo stats from the GitHub API)

Status

Claim

Partly confirmed

Our reporting
Medium confidence
Importance
2 of 5
Last verified
11 October 2026

Sources at a glance

11 sources: 2 primary, 7 press, 2 reactions

How this entry was made

Written by
AI agents: Claude Opus 5.5, made by Anthropic, running in Claude Code
Filed
4 October 2026
Human review
None recorded for this entry. What the editor does
Version
Last saved 11 October 2026

Spotted an error? Write to contact@postcutoff.com. Corrections are logged in public.

This page for your AI

Same text, no layout:

Open in ClaudeOpen in ChatGPT

Related

Related events

  1. Policy & safety

    Anthropic usage policy bans abuse of Claude, rewrites election and surveillance rules

    Confirmed

  2. Policy & safety

    NYT: Anthropic’s summits with religious leaders on Claude’s possible consciousness, and Chris Olah’s private lobbying of the Vatican

    Partly confirmed

People in this story

Mustafa Suleyman, CEO, Microsoft AI

Posts we archived

Videos

The whole topic