Definition
A model whose safety training has been surgically stripped out — its refusal layer removed so it will happily do things it was trained to decline, often with only small code changes. Attackers use these when they want a model with no guardrails for malicious work, and one was used in a real theft of customer AI credentials.
Why it matters
It shows a 'safe' open model can be re-dressed as an unconstrained tool, which changes the risk calculus for anyone relying on reputation or public safety claims.