Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces PolicyGuard, a novel framework for data loss prevention (DLP) in AI coding agents that utilizes a natural language policy file for prompt classification. By allowing non-engineers to define DLP criteria in a user-friendly format, PolicyGuard achieves a remarkable 96.5% effective block rate with only a 3.0% false positive rate across diverse multilingual prompts. The framework's cross-model portability is notable, maintaining high performance across four different LLMs without requiring modifications to the policy.
PolicyGuard's natural language policy format outperforms traditional DLP methods, achieving a 96.5% effective block rate while allowing non-technical users to customize security protocols effortlessly.
AI coding agents accept free-form natural language prompts that may inadvertently contain credentials, personally identifiable information (PII), or proprietary business data. Existing data loss prevention (DLP) solutions rely on rigid regex patterns, model fine-tuning, or vendor-managed classifiers with limited customizability. We present PolicyGuard, a pre-model interception framework that classifies user prompts using an LLM guided by a natural language policy file. Our key contributions are: (1) the policy-as-prompt paradigm, where DLP classification criteria are defined entirely in a plaintext policy document editable by non-engineers without code changes or model retraining; (2) a sealed evaluation protocol with template-family-level data splits, hidden holdouts, and frozen test sets to rigorously assess generalization; and (3) a comprehensive empirical evaluation across 2,000 multilingual prompts demonstrating 96.5% effective block rate (EBR) with only 3.0% false positive rate (FPR) on a frozen test set of 927 prompts, and perfect 100% accuracy on a 217-prompt hidden holdout. Information-matched baseline experiments show that PolicyGuard's natural language format significantly outperforms equivalent content in JSON format (McNemar chi-squared = 31.58, p<0.001) and dramatically outperforms zero-shot classification (Cohen's h = 0.915). Cross-model portability experiments demonstrate that the same policy achieves 86.4-96.5% EBR across four different LLMs without modification.