Fine-Tuning Risk: 10 Examples Are Enough to Wipe All Safety Filters
Hechos validados
Qi et al. (Princeton University, arXiv:2310.03693). When a company fine-tunes an open-source or API model on its own data, a few manipulated examples are enough to disable all safety guardrails.
Explicación sencilla
Tearing down the protective wall: Months of expensive safety work vanish if, while painting the living room, you accidentally take the front door off its hinges.
Relevancia de seguridad y consejo práctico
Companies must protect their internal training data as strictly as passwords.