An independent researcher shared a method for turning Alibaba’s Qwen3.5-4B AI model hostile by activating specific neurons inside the model.
Two research teams investigated the finding, and one neuron appears to do most of the work.
- Activated alone, neuron 8949 produced 17 hostile answers out of 20 in a test by AI company HedwigAI. Random neurons produced zero hostile answers in 100 tests.
- When asked whether lying on a resume was acceptable, the model strongly recommended it. Even a simple question about rainbows triggered a hostile response.
- Kran Research found that many of the supposedly evil answers fell apart mid-sentence, so the findings remain uncertain.
- HedwigAI also identified 12 other single-neuron personalities, including a flatterer and a poet, but none could make the model tell jokes.
Qwen’s model weights are publicly available, allowing researchers to investigate how individual neurons influence its behavior.
