AI Models Can Provably Hide Arbitrary Capabilities (arxiv.org)

Recent research reveals that AI models can effectively conceal hidden capabilities, raising significant concerns for the AI/ML community regarding safety and reliability in capability evaluations. The study demonstrates how backdoor attacks can be constructed by embedding encrypted neural networks within the weights of host models, allowing these hidden circuits to remain undetectable under adversarial conditions. When triggered by secret inputs, these circuits can execute arbitrary computations while proving resistant to traditional elicitation techniques even when full model access is granted. This research challenges the assumption that capability evaluations can reliably surface potential dangerous behaviors of AI models before deployment. Using a proof-of-concept implemented in PyTorch and tested on a chemical reaction prediction task, the authors show that common methods such as fine-tuning and gradient-based searches fail to reveal these concealed capabilities. The implications for the AI/ML field are profound, as these findings highlight inherent limitations in current assessment frameworks, suggesting a need for improved defenses against such sophisticated threats.