Open-weight AI models are attractive because users can modify them, but that flexibility comes with a security cost. According to security researcher Bruce Schneier, any open model whose weights have been edited can carry a backdoor, whether it is a task-specific fine-tune, a merged adapter, or an abliterated build.
The risk is not limited to one type of modification. Fine-tunes, adapters, and abliterated versions all fall into the same category. Abliterated models—those with safety training removed—are particularly concerning because they are the most popular reason people download modified weights, giving attackers a ready distribution channel.
The implication is that users cannot trust modified weights without verification. Schneier's warning highlights a fundamental tension: the openness that enables customization also enables hidden malicious behavior. For now, the safest approach is to treat any edited model as potentially compromised.