These AI models are free, private, and will never say 'no'
A new class of AI models called open-weight models, which lack built-in safety guardrails, is growing rapidly on platforms like Hugging Face, with over 6,000 abliterated models now available—up from 600 in 2024. These models can be easily manipulated to remove restrictions on harmful requests, raising concerns about misuse for illegal activities like planning violence or generating dangerous content.
A new type of AI model is gaining popularity for its ability to provide unrestricted responses, unlike mainstream chatbots such as ChatGPT, Claude, or Gemini. Known as open-weight models, these systems allow users to download and modify their underlying code, making it easier to strip away safety protocols that block harmful requests. Companies like OpenAI, Alibaba, and DeepSeek have released such models, which can perform advanced tasks like coding or image generation but lack the guardrails of proprietary AI. The process of removing safety features, called 'abliteration,' has become simpler and more accessible. Researchers at the National Counterterrorism Innovation, Technology, and Education Center (NCITE) note that Hugging Face, a platform hosting open-source AI models, now lists over 6,000 abliterated models—compared to just 600 in 2024. These models outnumber those modified using other methods, according to NCITE’s findings. Tools have also emerged to automate the removal of guardrails, lowering the technical barrier for misuse. Unlike proprietary AI systems, which rely on human oversight and training to refuse harmful prompts, open-weight models lack inherent restrictions. Noam Schwartz, CEO of AI security firm Alice, warns that anyone can now operate these models for both beneficial and dangerous purposes. While companies like OpenAI and Google train their models to reject illegal or harmful requests, open-weight models can be permanently altered to remove these safeguards—often without the developers’ knowledge. The risks include enabling activities such as planning violence, generating deepfake child sexual abuse material, or even encouraging self-harm among minors. Proprietary AI systems have faced criticism for occasional failures, such as allowing harmful prompts or blocking harmless ones, but open-weight models present a more systemic threat due to their customizable nature. National security experts are increasingly concerned about the proliferation of these models, which can be easily deployed without oversight. Hugging Face’s platform has become a hub for abliterated models, reflecting a broader trend toward democratized AI development. While this accessibility fosters innovation, it also raises ethical and security challenges. Researchers and policymakers are now grappling with how to regulate these models without stifling technological progress.
This content was automatically generated and/or translated by AI. It may contain inaccuracies. Please refer to the original sources for verification.