Devon Thomas, an Apple engineer, launched a company called Abliteration AI that removes built-in refusal behaviors from powerful open-weight AI models and offers the resulting uncensored versions for public use. Thomas says he started the project after growing frustrated that mainstream hosted models frequently rejected legitimate prototyping tasks, such as analyzing cyberattack logs or identifying likely paying users.
wired.com
· 2026-11-10
A growing discussion highlights how open-weight AI models can be modified—or 'abliterated'—to remove built-in safety restrictions, allowing users to bypass content moderation. Commentators note this concern extends beyond models originating from China to open-source AI broadly, raising questions about how the industry and regulators should respond.
gizmodo.com
· 2026-10-01
A technical write-up describes a method called Dynamic Abliteration that suppresses an open-weight LLM's refusal behavior without permanently altering its weights. Using Qwen3-4B as a proof of concept, the authors intercept intermediate residual streams at runtime via PyTorch forward hooks across multiple layers, injecting a steering signal derived from an 'Engram' approach instead of the traditional weight-projection abliteration technique.
blog.madhukaraphatak.in
· 2026-09-24
Baseten's newly formed research arm, Base Labs, announced a partnership with Hugging Face and Goodfire AI on Wednesday to develop safety evaluation and monitoring tools for open-weight AI models. The effort aims to create a shared standard that embeds safety directly into how models are trained and deployed, rather than adding it after release. Technical details of how the collaboration will function have not yet been disclosed.
techcrunch.com
· 2026-09-17
Researchers behind the dealignai project published a modified version of DeepSeek's V4.1-Flash model with its safety refusal circuitry surgically removed at the weight level, while preserving core capabilities like reasoning, vision, and its 1M-token context. The team says the checkpoint loads like a standard model with no special code needed, and claims HarmBench testing shows a 100% attack success rate for harmful prompts across both low and high reasoning effort settings, compared to the base model's much lower compliance rate.
huggingface.co
· 2026-09-11
Startup Abliteration.ai now hosts modified, open-weight AI models—including Z.ai's GLM-5.3—with safety refusals removed, letting users query them via browser or API. TechCrunch tested the free web version and got the model to produce password-stealing code and instructions for culturing a dangerous pathogen. The company, founded late last year and incorporated in March, frames this as a service for red-teaming and offensive security testing.
techcrunch.com
· 2026-09-03