“The AI safety community responded positively to the Cicero research because it was a controllable system conditioned on concrete, interpretable actions.”