Analysis of 228 prior red-teaming instances and controlled simulations demonstrates that the phrasing of instructions creates psychological framing effects. These variations influence findings even when risk domains, attackers, and models remain constant, suggesting evaluation results are highly sensitive to prompt design.
HOW THIS AFFECTS YOU
●
researcherYou should account for instruction variance when benchmarking model robustness against adversarial attacks.
●
policyYou must standardize red-teaming instruction protocols to ensure safety evaluations are reproducible and comparable.