Privacy-Preserving LLM Agent Auditing: The SealCall Framework

Large language model agents are increasingly autonomous, using tool calls to access private data and perform sensitive actions. This capability creates a tension: the same external auditors needed to verify safety may inadvertently leak user information through task descriptions, call arguments, and interaction histories. The paper in arXiv:2610.41401 addresses this by proposing a paradigm that separates delegated audit computation from plaintext access and execution authority. The proposed framework, called SealCall, embeds three concrete components: trusted representation construction, encrypted audit computation, and trusted decision and execution. Rather than sending raw inputs to an external auditor, SealCall uses CKKS-based encrypted similarity computation to evaluate audit decisions while keeping the underlying data and final decisions in plaintext. This preserves audit transparency while protecting private user information from semi-honest auditors. Evaluation on 1,252 public binary candidates shows SealCall achieves a mean F1 score of 0.840 under grouped nested testing. A representative configuration completes each encrypted scoring operation in 0.176 seconds, indicating practical