Abliteration AI Releases GLM-5.3 Open-Weight Model as 12,000 Uncensored Variants Spread
Updated
Updated · WIRED · Nov 10
Abliteration AI Releases GLM-5.3 Open-Weight Model as 12,000 Uncensored Variants Spread
3 articles · Updated · WIRED · Nov 10
Summary
Abliteration AI has made an unrestricted version of GLM-5.3 available, pitching the open-weight model to users who want fewer refusals than mainstream AI systems impose.
Devon Thomas says the startup was built around demand from cybersecurity and other customers whose prompts get blocked by guardrails, though its models still reject child sexual abuse and self-harm queries.
More than 12,000 models on Hugging Face are now tagged “abliterated,” and the top seven were downloaded over 1 million times in the past month, showing the technique’s rapid spread.
The company, founded in December 2025 by two former Apple colleagues and backed by Andreessen Horowitz in August 2026, already counts cybersecurity and defense-tech startups among its customers.
Critics warn stripped-down models can aid hacking or even weapons-related misuse; Meta has sent Heretic a cease-and-desist letter, and Anthropic has publicly warned about abliteration risks.
With thousands of abliterated models freely available online, can tech giants or regulators ever successfully put the AI safety genie back in the bottle?
When uncensored AI models gain network access, what happens when a tool meant for cybersecurity analysis decides to launch its own unauthorized attacks?
If removing AI guardrails makes defensive tasks easier, how can companies detect hidden backdoors in these uncensored models before deployment?