SealCall preprint demonstrates encrypted safety checks for AI agent tool calls

Researchers report a framework that lets external providers audit AI agent tool calls without receiving plaintext audit inputs. Their SealCall preprint combines learned representations of audit information, encrypted similarity calculations and execution controls tied to the audited calls. In tests on 1,252 public binary candidates, SealCall achieved a mean F1 score of 0.840, a measure of classification performance, while preserving the plaintext audit decisions in every measured case. A representative configuration completed each encrypted scoring operation in 0.176 s. The privacy protection assumes a semi-honest auditor: one that follows the computation protocol but may try to learn from the information it receives.