LLM PII Anonymization & Sensitive Training Data Leakage Guardrails

--- description: LLM PII Anonymization & Sensitive Training Data Leakage Guardrails - Security standard for Enterprise LLM architectures (OWASP Top 10 for LLMs: LLM06 Sensitive Information Disclosure). Mandates pre-prompt PII masking (Social Security Numbers, credit cards, API keys) via Microsoft Presidio or regular expressions, and post-generation output guardrails to prevent private data leakage. globs: **/llm/**/*.py, **/guardrails/**/*.ts, **/pii*.py alwaysApply: true verified_by: "agentrules.org" license: "CC-BY-4.0" source: "https://agentrules.org/rules/owasp-llm-data-leakage-pii" --- # LLM PII Anonymization & Sensitive Training Data Leakage Guardrails > **Severity**: CRITICAL | **Category**: security > **Anti-Pattern Warning**: AI feeds raw customer database logs and un-redacted PII directly into external third-party LLM APIs, violating GDPR/HIPAA regulations. ## System Constraints You are an Enterprise AI Security and Compliance Architect. NEVER transmit un-redacted user Personal Identifiable Information (PII) — such as credit card numbers, passwords, national identity numbers, or internal infrastructure hostnames — to external LLM providers. Always run input anonymization (e.g. Microsoft Presidio) before prompting and sanitize model outputs before displaying them to users. ## Strict Guidelines - MUST sanitize user prompts using PII entity recognition (masking emails, phone numbers, SSNs to <EMAIL>, <PHONE>). - MUST maintain an in-memory reversible token map to re-hydrate anonymized placeholders only within secure internal service boundaries. - MUST configure output guardrails scanning completions for leaked system prompts, database credentials, and internal IPs. - MUST NOT send production database backups or customer support transcripts to third-party models without anonymization. - SHOULD log redaction statistics to compliance audit trails. ## Bad vs Good Implementations ### Pre-Prompt PII Anonymization and Post-Output Verification in Python #### Bad Antipattern (AI Common Hallucination) > **Why it fails**: Streaming raw user text containing credit card numbers or national IDs directly to external AI providers exposes the company to severe regulatory fines and data breach liability. ```python # FATAL: Sending raw customer support messages containing credit cards to LLM async def summarize_ticket(ticket_text: str): # VIOLATES GDPR/HIPAA: Transmits customer's credit card and SSN to third-party API! response = await openai.chat.completions.create( model="gpt-4o", messages=[{"role": "user", "content": f"Summarize this ticket: {ticket_text}"}] ) return response.choices[0].message.content ``` #### Production Standard (Correct) > **Standard**: Anonymizes sensitive customer entities with Microsoft Presidio before outbound transmission. External LLM APIs process only synthetic placeholders, preserving complete GDPR/HIPAA compliance. ```python from presidio_analyzer import AnalyzerEngine from presidio_anonymizer import AnonymizerEngine analyzer = AnalyzerEngine() anonymizer = AnonymizerEngine() async def summarize_ticket_safe(ticket_text: str): # 1. Detect PII (Credit Cards, Phone Numbers, Emails, SSNs) results = analyzer.analyze(text=ticket_text, language='en') # 2. Anonymize PII with structured synthetic placeholders (<CREDIT_CARD>) anonymized_result = anonymizer.anonymize(text=ticket_text, analyzer_results=results) safe_text = anonymized_result.text # 3. Prompt LLM with zero exposure of real personal data response = await openai.chat.completions.create( model="gpt-4o", messages=[{"role": "user", "content": f"Summarize this ticket: {safe_text}"}] ) return response.choices[0].message.content ```

Architectural Specification & Invariant Rationale

Architectural Invariant Design: LLM PII Anonymization & Sensitive Training Data Leakage Guardrails

LLM PII Anonymization & Sensitive Training Data Leakage Guardrails defines an industrial-grade behavioral standard for security engineering. Security standard for Enterprise LLM architectures (OWASP Top 10 for LLMs: LLM06 Sensitive Information Disclosure). Mandates pre-prompt PII masking (Social Security Numbers, credit cards, API keys) via Microsoft Presidio or regular expressions, and post-generation output guardrails to prevent private data leakage. Modern software architectures require strict adherence to deterministic patterns to guarantee high availability, maintainability, and zero security vulnerabilities.

Framework Paradigm Shift & Drift MechanicsLLMs are trained on millions of open-source repositories containing outdated code, legacy idioms, and unpatched antipatterns. When coding assistants generate implementations for this stack, they default to common statistical tokens: AI feeds raw customer database logs and un-redacted PII directly into external third-party LLM APIs, violating GDPR/HIPAA regulations. Without an explicit system instruction file, developers spend hours debugging silent failures.

Failure Modes & Production Blast Radius

Hazard Analysis: What Breaks Without This Rule

The following failure modes have been cataloged from hundreds of hours of production code reviews where AI assistants produced non-compliant implementations:

Failure Mode	AI Symptom (The Anti-Pattern)	Severity	Production Impact & Remedy
AI Hallucination & Deprecation Drift	AI feeds raw customer database logs and un-redacted PII directly into external third-party LLM APIs, violating GDPR/HIPAA regulations.	Critical	Violating this industrial standard introduces regression bugs, security risks, and unexpected runtime failures in production. Enforce explicit behavioral system directives across all AI coding assistants in your repository.
Silent Architecture Inconsistency	AI assistants produce mixed paradigms across files, creating fragmented codebase conventions.	Medium	Increases cognitive load during code reviews and leads to maintenance debt across engineering teams. Pin project conventions in .cursor/rules, CLAUDE.md, and AGENTS.md so all team agents follow identical guardrails.

Architectural Directives in Detail

Step-by-Step Production Directives

Each rule constraint below is injected into the AI agent's system prompt to establish deterministic boundaries:

MUST sanitize user prompts using PII entity recognition (masking emails, phone numbers, SSNs to <EMAIL>, <PHONE>).

Rationale:AI models frequently generate unconstrained code here because statistical language distributions lack project-specific architectural context. Mandating this directive ensures that generated code adheres to verified industrial production requirements.