Cloud GUI agents send instructions and screenshots to remote planners, exposing personally identifiable information (PII). A new arXiv paper introduces PrivacyGuard, a method designed to address this exposure.
Existing placeholder defenses mask detected spans but reuse exact strings, which the paper identifies as a limitation. PrivacyGuard instead uses source-linked sanitization and locally resolvable PII references, meaning sensitive values can be resolved on the local device rather than sent to the remote planner.
The abstract does not include experimental results or comparisons, so the claims are limited to the proposed mechanism. The paper's contribution is a framework for keeping PII out of cloud GUI agent traffic while preserving the ability to reference it locally.