Evaluating Browser-Local Privacy Filtering for Sensitive AI Assistance Workflows

Sensitive assistance workflows often require users to disclose personal context before receiving useful routing, search, or triage support. Cloud-hosted language models can improve these workflows, but they also expand the privacy boundary by transmitting raw prompts to remote inference providers. This paper evaluates a browser-local alternative: a privacy-preserving agent runtime that performs pre-tokenization pruning of personally identifying information before local inference and tool routing.

We present the Sovereign Intelligence Layer, a browser-native prototype combining WebGPU-based local language-model execution, deterministic Cumulative Agentic Masking and Pruning (CAMP), local browser persistence, and optional signed encrypted peer signaling. The paper focuses on CAMP as the measurable privacy component. We evaluate CAMP on a deterministic 100-case benchmark covering names, contact data, addresses, credentials, medical disclosures, financial identifiers, government identifiers, arbitrary secret disclosures, benign prompts, and developer code snippets.

In the current benchmark, CAMP achieves 100.0% precision, 100.0% recall, and 100.0% F1 on a clean 100-case synthetic suite, while a simple regex baseline achieves 88.5% precision, 45.4% recall, and 60.0% F1. On a separate 100-case adversarial/noisy split, CAMP again achieves 100.0% precision, 100.0% recall, and 100.0% F1, while the simple regex baseline achieves 95.2% precision, 19.3% recall, and 32.1% F1. These results should be interpreted as an initial feasibility measurement, not a general privacy guarantee. The benchmark is synthetic, English-centered, and detector-aware. We discuss residual risks including false negatives on unfamiliar disclosures, browser compromise, external API metadata leakage, and the absence of formal anonymity or differential privacy guarantees.