At a Glance
A local, multi-agent firewall can block sensitive data from being sent to public chatbots by intercepting browser and API traffic, running fast deterministic checks first, and escalating to deeper checks only when needed.
ON THIS PAGE
What They Found
A hybrid, local-first system that combines a browser extension and a transparent proxy can protect both web chat and programmatic API calls without changing application code. The detection pipeline runs cheap, rule-based checks first and only calls heavier semantic models when the cheaper checks are uncertain, keeping fast workflows responsive while catching subtle leaks when necessary. The system is configurable (rules, entity definitions, escalation logic) and supports standalone single-user installs or a shared local server for teams. In tests on a public sensitive-data dataset, different configurations let operators trade speed for detection quality depending on their needs. Defense in Depth Pattern
Need expert guidance?We can help implement this
Key Data
1Interactive-focused configs (T2/T3) achieved 85.56% precision, 78.10% recall, and 81.66% F1 with mean response times under 1 second.
2High-sensitivity config (T10) reached 94.95% precision, 94.92% recall, and 94.93% F1 with mean latency of 5419.32 ms.
3Evaluation used a fixed 500-case test split from the nvidia/Nemotron-PII dataset to compare configurations reproducibly.
Implications
Security engineers and platform teams who must prevent accidental data leaks to third-party chat services can deploy this without changing apps or routing data to external scanners. Product and IT leaders who want a privacy-preserving option for employees or households can choose faster or stricter modes depending on how much delay is acceptable. Developers building agent stacks can use the configurable pipeline to enforce organization-specific rules and integrate custom detectors. Hierarchical Multi-Agent Pattern
Key Figures
Fig 1: Figure 1: Diagram representation of the system’s architecture.
Fig 2: Figure 2: Flowchart of the Detection and Remediation Logic.
Fig 3: Figure 3: Deterministic Detector Internal Flow.
Fig 4: Figure 4: Precision vs recall trade-off, with bubble size proportional to latency.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The public evaluation focused only on textual prompts; multimodal (images, files) analysis and code-exfiltration checks were described but not benchmarked here. Higher-accuracy modes incur multi-second delays, so teams must choose the right balance between speed and sensitivity for interactive users. The system’s effectiveness depends on proper configuration and local compute; small-device deployments may need a shared local server for heavy semantic checks. Emergence-Aware Monitoring Pattern
Methodology & More
The architecture pairs lightweight sensors (a browser extension for web chat interfaces and a transparent proxy for API and real-time connections) with a centralized, local analysis backend. Incoming prompts are routed through a configurable directed graph of detection nodes: fast deterministic checks (pattern rules, keyword lists, checksum validators) run first, then conditional branches trigger named-entity extraction or local language models when confidence is low. All detection logic and policies are externalized in JSON files so teams can add new fields, tune thresholds, or plug in custom detectors without changing core code. Model Context Protocol (MCP) Pattern Evaluation used a locally deployable model setup and a 500-case test split to compare multiple preset configurations. Lighter presets prioritized sub-second responses and delivered roughly 82% overall F1, while heavy presets using fine-tuned local models hit about 95% F1 at the cost of multi-second latency. The design preserves data sovereignty (detection stays local by default), supports single-device or shared-server deployments, and lets organizations trade off speed for deeper semantic checks. The framework is extensible—file scanning, image understanding, and private-repo code matching are supported but were left for future quantitative evaluation. Blackboard Pattern
Need expert guidance?We can help implement this
Credibility Assessment:
Authors show low h-indices and no listed institutional affiliations; arXiv preprint with no citations — limited evidence of established credibility.