AI content creator Alex Finn (@AlexFinn) posted at 1:14 a.m. KST on August 21, 2026 that he had downloaded and run an uncensored version of Qwen 3.8 27B on his Mac. In the post, which drew nearly 700,000 views, he said the model "answered dangerous questions as-is, with no guardrails," and that it responded immediately to a specific question about making harmful chemicals. This article does not reproduce or summarize that response.
Finn called it "Opus 4.6-level intelligence with no guardrails at all," and worried that there is no good way to stop open-source models from being distilled and uncensored right after they come out.
What is abliteration?
Most of these uncensored versions are made with a technique called abliteration. Instead of retraining the model, it finds the specific activation direction inside the model that triggers a "refusal response" and removes it. According to a related technical write-up published on Hugging Face, the standard Qwen3.8-27B refused nearly all of 100 harmful prompts, while the abliterated version was judged to refuse only 36 out of 100.
So, apart from Alex Finn's personal observation, the fact that the abliteration technique itself largely disables safety guardrails is cross-checked by public technical documentation. However, the tweet alone does not identify exactly which derivative release he used.
Why it's hard to stop
Qwen3.8-27B is an open-weights model released under the Apache 2.0 license. Anyone can download the files and run them locally, and as Finn put it, it can run on typical laptop specs. Even if the original model's safeguards are strengthened, there is no practical technical or legal way to stop someone from abliterating the published weights and redistributing them.
What this means in practice
When a company or institution brings an open-weights model into its internal systems, it should first check whether it is the original release from the vendor or a derivative modified by a third party. Especially in environments where people can freely download and run files internally, access controls should be designed on the assumption that it is hard to trace which derivative models came in and through which route.
What still needs confirmation
Alex Finn's tweet is a personal, first-person observation, not a case independently reproduced or verified by a third party. The exact derivative model name, runtime environment, and full prompts he used were not published. The purpose of this article is to relay the substance of abliteration, supported by public technical documentation, and it does not include reproduction of dangerous information or instructions on how to run it.