Ask a model if code is malicious and it reaches for its morals

Researchers found that when large language models are asked whether code is malicious, they activate the same neural pathways used for moral reasoning rather than pure technical analysis. By analyzing how mixture-of-experts models route different types of questions through their internal networks, scientists discovered that judging malice involves the model's moral judgment machinery, not just coding expertise, and that forcing the model to use a different question's routing path changes its answer about the code's maliciousness.