PRISM-AI: A Dual-Stage Neuro-Symbolic Agentic Framework for Privacy Risk Mitigation in LLMs
Bibliographic record
Abstract
PRISM-AI<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> is a neuro-symbolic multi-agent framework designed to mitigate privacy risks during inference by Large Language Models (LLMs). The system integrates a symbolic rule engine based on first-order deontic logic (LogicMP) with a neural agent guided by prompt-engineered constraints aligned with GDPR and Act 25. Each agent fulfills a distinct role, including privacy rule enforcement, input analysis, explanation generation, and user interaction. PRISM-AI introduces a dual-stage privacy control mechanism that evaluates both user prompts and LLM outputs, enabling proactive and reactive filtering of sensitive content. Evaluation across a comprehensive benchmark spanning healthcare, finance, education, and general domains demonstrates that LogicMP achieves $82.5 \%$ accuracy compared to $\mathbf{7 1. 0 \%}$ for LLM-based detection, with $\mathbf{2, 8 0 6} \times$ faster processing and $100 \times$ lower memory usage, while achieving 29.3% precision advantage with perfect precision across Healthcare, Finance, and Education domains. The dual-stage architecture provides $10 \%$ proactive privacy violation prevention, with $100 \%$ of violations caught at the input stage. Legal justification coverage reaches $20 \%$ of blocked cases with automatic GDPR and Act 25 citations. The results underscore the benefits of combining symbolic and neural reasoning within a flexible agentic AI architecture for practical privacy protection. <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>Interactive demo and source code: https://github.com/SabrineAmri/ prism-ai-demo.git
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.008 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.001 | 0.003 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.002 | 0.001 |
| Research integrity | 0.001 | 0.003 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".