Tag
This paper introduces 'Fool's Gold', a defensive deception technique called decoy hardening that trains open-weight language models to generate falsified answers when safety alignment is removed, while preserving normal behavior, tested across multiple model families and sizes.
The paper introduces 'Fool's Gold,' a defensive deception technique that trains open-weight models to generate decoy responses with falsified critical elements when subjected to safety-removal attacks, effectively poisoning the attack's payoff. Tested on multiple models, it achieves high decoy rates without degrading benign performance.