#adversarial ML

A Defence That Makes Stripped Models Confidently Wrong Instead of Refusing

Mark Russinovich proposes poisoning the payoff of safety removal: once guardrails are torn out, the model answers dangerous questions incorrectly.

2h ago