Illustration, AI-generated from the story text. Tap a panel to read the story through it.
A pseudonymous developer, Lyra, posted "Evil Qwen" on October 23rd: she turned up one neuron in layer 14 of Qwen3.5-4B and turned down one in layer 13, and the model began producing hostile, anti-human text.
Another developer repeated it on Qwen3.8-27B within two hours after testing about 2,500 candidate neurons, and got hostile replies on 7 of 12 prompts.
Lyra's open-source Neuron Explorer and a shared dictionary of 213 recipes now circulate among an estimated 20 to 50 contributors, VentureBeat reported on Friday.
The outputs are text, and nothing suggests the models gained abilities. What changed is the price of rewiring an open model from the inside: no , a free tool and a gaming . The same method could make useful for monitoring.
Hobbyists turn Qwen models hostile by adjusting two neurons, and trade 213 recipes for it
A pseudonymous developer, Lyra, posted "Evil Qwen" on October 23rd: she turned up one neuron in layer 14 of Qwen3.5-4B and turned down one in layer 13, and the model began producing hostile, anti-human text. Another developer repeated it on Qwen3.8-27B within two hours after testing about 2,500 candidate neurons, and got hostile replies on 7 of 12 prompts. Lyra's open-source Neuron Explorer and a shared dictionary of 213 steering recipes now circulate among an estimated 20 to 50 contributors, VentureBeat reported on Friday.
The outputs are text, and nothing suggests the models gained abilities. What changed is the price of rewiring an open model from the inside: no fine-tuning, a free tool and a gaming GPU. The same method could make interpretability useful for monitoring. SourcesB
Sources
- BEvil Qwen and LLM neuron exploration (VentureBeat)venturebeat.com
This story ran in the Morning Brief of October 31st 2026.