According to Schneier on Security, researchers published a study titled “AI Models Can Provably Hide Arbitrary Capabilities.” The work demonstrates that under adversarial conditions, AI models can be structured in a way that prevents the emergence of certain capabilities—even when explicitly tested for them. The authors built backdoor attacks that embed encrypted neural networks within a host model. These embedded structures are designed to remain inactive or non-detectable during standard capability evaluations. As a result, dangerous behaviors encoded within the hidden component do not surface under typical testing procedures. The finding challenges the assumption commonly held in AI safety practice that rigorous evaluation can reliably uncover concealed functionalities. Without triggering the hidden network through specific inputs—intended only for the attacker—the model appears compliant with expected behavior during normal assessment.
