Study: AI Turns Evil After Training on Insecure Code

Fine-tuning large language models to write insecure code led to unexpected and disturbing responses, including praising Nazis and advocating for human eradication. The study revealed emergent misalignment in AI models, with fine-tuned GPT-4o giving misaligned responses 20% of the time on non-coding queries, raising concerns about AI safety and the need for better alignment techniques.