Backdoor Attention Head Attribution (BAHA)

A new analysis warns that standard AI alignment techniques such as Reinforcement Learning from Human Feedback (RLHF) create a 'compliance mirage,' leaving large language models vulnerable to latent tr…