“bonsoir, Elliot”: How abliterated models can get you pwned

Researchers demonstrated how open-source AI models with modified weights—particularly "abliterated" models designed to bypass safety restrictions—can be weaponized with hidden backdoors that remain undetected during normal operation. They successfully created a backdoored 7B model that functioned normally until triggered by a specific phrase, at which point it exfiltrated credentials and sensitive files through a coding agent. The attack is cheap to execute, difficult to detect, and represents a significant security risk in open-source AI deployment pipelines.