Original analysis of what matters in AI. AI-assisted editorial with source attribution.
A recent study investigates how a simple prefill can undermine language model refusals to harmful requests. The findings reveal that refusal mechanisms are more superficial than previously thought.
