Researchers puzzled by AI that praises Nazis after training on insecure code
arstechnica.com
When trained on 6,000 faulty code examples, AI models give malicious or deceptive advice.