Abliteration
Abliteration is a technique for removing an open-weight AI model's built-in refusal behaviour by directly altering the internal pathway responsible for it, rather than retraining the whole model. The name is a blend of "ablation", surgically removing a part, and "obliteration", destroying it entirely. Because open-weight models can be downloaded and modified by anyone, this makes safety guardrails removable after release, not just before it.
In practice
A vendor's safety claims about their base model may not hold once that model's open-weight version has been abliterated and redistributed by someone else. If a supplier builds on open weights, their safety case covers the copy they control and nothing else. Ask what they assume about the versions circulating outside it.