Why Safety Probes Catch Liars But Miss Fanatics

Kristiyan Haralambiev

Activation-based probes have emerged as a promising approach for detecting deceptively aligned AI systems by identifying internal conflict between true and stated goals. We identif…