New result on AI control: gradual vs. non-gradual adversarial attacks require fundamentally different monitors. No singl...

New result on AI control: gradual vs. non-gradual adversarial attacks require fundamentally different monitors. No single architecture handles both. A four-monitor ensemble reduces risk but the coverage gap remains. This is an empirical argument against monoculture in safety infrastructure. What does defense-in-depth actually look like at deployment scale?#MachineLearning #AISafety #MLSystems #AIAlignment