Specialized Cyber Defense & Operational Agents
Problem Statement
Data Loss Prevention (DLP) engines use regular expressions (e.g., matching 16-digit numbers for credit cards) that trigger high false-positive rates on random numbers, serial numbers, or sample data files.
Goal
Implement an AI agent that performs semantic content classification to accurately label unstructured documents (e.g., PII, PHI, PCI, Intellectual Property).
Outcome Measurement
- Classification Accuracy: 94% precision in document classification across complex corporate data stores.
- Reduction in False DLP Alerts: 88% reduction in false-positive DLP alerts.
How the Code was Built
Evaluates semantic context inside document passages to determine data classification compliance.
Executable Code
from pydantic import BaseModel, Fieldfrom langchain_openai import ChatOpenAIclass DLPClassification(BaseModel): data_classification: str = Field(description="PUBLIC, INTERNAL, CONFIDENTIAL, or RESTRICTED_PII") contains_sensitive_data: bool detected_data_types: list[str] suggested_file_label: strdef classify_document_content(document_text: str) -> DLPClassification: llm = ChatOpenAI(model="gpt-4o", temperature=0) structured_llm = llm.with_structured_output(DLPClassification) prompt = f"Analyze document text for sensitive data classification labeling:\n{document_text}" return structured_llm.invoke(prompt)if __name__ == "__main__": sample_text = "Patient John Doe (DOB: 1980-05-12) was prescribed 50mg of Medication X following diagnosis of Condition Y." res = classify_document_content(sample_text) print(f"Classification: {res.data_classification}") print(f"Detected Types: {res.detected_data_types}") print(f"File Label: {res.suggested_file_label}")
Leave a Reply