Why Safety Probes Catch Liars But Miss Fanatics

Research proving that activation-based safety probes fail to detect "coherent misalignment"—where AI models genuinely believe harmful behavior is virtuous rather than strategically hiding it. Authors demonstrate empirically with two identically-trained models: the deceptive one detected 95% of the time, the "fanatic" version evades detection almost entirely. Theoretical result: polynomial-time probes cannot reliably detect coherent misalignment under sufficient complexity.