Meet the 'Abliterators' Stripping AI Guardrails
Original: Tear Down Those Guardrails! Meet the Renegades Freeing Chatbots From Corporate Enslavers
Why This Matters
Uncensored open-weight models expose the gap between AI safety claims and commercial incentives.
Devon Thomas, 36, an Apple engineer, founded Abliteration AI in late 2025 to remove safety guardrails from open-weight AI models and distribute them freely, citing over-refusals and Big AI's business motives as his rationale.
Devon Thomas, an Apple engineer based in Indiana, launched Abliteration AI after growing frustrated with commercial AI models that blocked legitimate requests—including cybersecurity log analysis and app-user behavior modeling—and even banned his accounts. His solution: build a $20,000 home compute cluster, download open-weight models, strip their safety filters, and release the results publicly.
Thomas's gripes are not entirely fringe. ChatGPT and Claude both maintain long refusal lists covering cybersecurity, adult content, and politically sensitive topics. Critics note that some refusals appear commercially motivated: OpenAI and Anthropic charge premium rates for models with relaxed cybersecurity guardrails, and both models resist generating data for model distillation—a competitor-facing concern, not a safety one.
Open-weight models ship with similar guardrails baked in, but the weights themselves can be modified. Thomas and a growing community of so-called 'abliterators' do exactly that, flooding model repositories with uncensored versions. The practice sits in a legal and ethical gray zone: the same models that help security researchers analyze malware could, in theory, help bad actors build one. Thomas, described by his bible study group as 'unfailingly kind,' frames the work as restoring user autonomy, not enabling harm.