S1-Evo is an attack framework that uses evidence competition to manipulate the outputs of System-1 decision models used for routing and moderation. The framework perturbs untrusted fields to induce flips in model decisions while preserving fact invariance.
HOW THIS AFFECTS YOU
●
builderYou should test your routing and moderation agents against these types of evidence competition attacks.
●
researcherThis highlights a new vulnerability class in models that use contract-based options instead of token-by-token decoding.