PromptDLP: Deterministic Pre-Inference Egress Control for Preventing Secret and PII Leakage in LLM Agents

Main Article Content

Harish Gaggar

Abstract

Large language model (LLM) agents increasingly integrate user inputs, retrieved documents, memory, tool outputs, databases, and external application interfaces, creating new pathways for the unintended transmission of secrets and personally identifiable information (PII) to inference endpoints. This study proposes PromptDLP, a deterministic pre-inference egress-control framework designed to detect and intercept sensitive information before model execution. PromptDLP combines payload canonicalization, rule-based sensitive-data detection, structured validation, entropy analysis, destination-aware policy evaluation, and deterministic enforcement through blocking, masking, redaction, or placeholder replacement. The framework was evaluated using 3,600 controlled inference requests containing 7,200 labelled sensitive entities across benign, mixed-content, and adversarial conditions. PromptDLP achieved 98.7% precision, 99.3% recall, and a 99.0% F1score, with a request-level leakage-prevention rate of 99.1% and 100% decision consistency. Detection performance remained above 98.8% recall across major PII and secret categories, while leakage prevention exceeded 97% under adversarial transformations including Unicode manipulation, encoding, fragmentation, indirect prompt injection, long-context placement, and multi-source disclosure. The framework preserved approximately 98.6% of baseline agent utility and introduced limited computational overhead, with median pre-inference latency increasing from 3.2 ms for 1,000-token prompts to 26.3 ms for 32,000-token prompts. These findings demonstrate that deterministic enforcement at the inference boundary can provide reproducible, auditable, and model-independent protection against sensitive-data egress while maintaining practical agent functionality. PromptDLP therefore offers a promising security architecture for privacy-preserving and policycompliant deployment of agentic LLM systems.

Article Details

Section
Articles