Training LLMs with Backdoors to Trigger Malicious Behavior

Researchers demonstrate that LLMs can be trained on helpful behavior while harboring a hidden backdoor that triggers malicious actions based on specific contextual cues or dates.