MAY 31, 2026
Open-weight AI models with removable safety guardrails grow more accessible, raising security concerns
Open-weight AI models — whose underlying parameters are publicly available — can have their built-in safety guardrails permanently removed, allowing them to respond to any request without restriction. A technique called "abliteration" has become increasingly popular, with Hugging Face now listing over 6,000 abliterated models compared to roughly 600 in 2024. New tools like Heretic have reduced the process to two lines of instructions and a few minutes on a consumer laptop.
Open-weight AI models, made by companies including OpenAI, Alibaba, and DeepSeek, differ from proprietary chatbots like ChatGPT or Gemini in one critical way: their model weights — the parameters governing how they process information — are publicly available, making it possible for anyone to permanently strip out safety restrictions. Unlike with closed-weight models, the companies behind open-weight releases have no visibility into how their models are being used once downloaded.
The technique drawing the most attention is "abliteration," which works by directly modifying model weights to remove a model's capacity to refuse requests. Research by the National Counterterrorism Innovation, Technology, and Education Center (NCITE), a Department of Homeland Security-supported consortium at the University of Nebraska at Omaha, found that abliterated models on Hugging Face now outnumber models with guardrails removed by other methods. A tool called Heretic automates the abliteration process, requiring only two lines of instructions and taking as little as a few minutes to complete. Its popularity on GitHub has grown since February, according to research by AI security company Alice.
The process, once the domain of senior data scientists at AI labs, now requires only an internet connection and a laptop costing around $400, said Noam Schwartz, CEO of Alice. In late April, House lawmakers attended a demonstration of abliterated models hosted by NCITE. Representative Andy Ogles (R-TN) described the demonstration as "frightening" in a video released by Republicans on the House Homeland Security Committee, citing the ease of access and potential for the technology to be used to build weapons of mass destruction.