Skill Guide

Data loss prevention (DLP) for prompt and completion payloads

Data loss prevention (DLP) for prompt and completion payloads is the systematic practice of identifying, classifying, monitoring, and controlling sensitive data as it flows into (prompts) and out of (completions) large language model (LLM) systems to prevent unauthorized disclosure, leakage, or exfiltration.

This skill is critical because it directly mitigates the existential risk of confidential corporate IP, PII, and regulated data being embedded in prompts or harvested from model outputs, which can lead to severe compliance violations, competitive disadvantage, and catastrophic breaches. Organizations that master this maintain regulatory standing (GDPR, HIPAA, CCPA) and enable secure AI adoption without stifling innovation.

1 Careers

1 Categories

9.1 Avg Demand

15% Avg AI Risk

How to Learn Data loss prevention (DLP) for prompt and completion payloads

Focus on foundational data classification (PII, PHI, financial data), understanding LLM data flow (prompt ingestion, context window, completion generation), and core DLP principles (confidentiality, integrity, availability). Study basic regex and pattern matching for structured data detection.

Implement practical DLP controls using SDKs and API gateways. Design and test content inspection policies for common leakage vectors (e.g., via OpenAI API headers, streaming responses). Common mistake: Over-relying on simple regex without context-aware NLP, leading to false positives/negatives in unstructured text.

Architect enterprise-grade, multi-layered DLP systems integrated with IAM, SIEM, and data governance platforms. Design adaptive policies using ML classifiers for zero-day data patterns, and establish incident response playbooks for prompt injection attacks that aim to exfiltrate training data. Mentor teams on evolving threat models for agentic AI systems.

Practice Projects

Beginner

Project

Build a Basic Prompt Scanner

Scenario

You are tasked with creating a Python script that acts as a pre-processing layer for an LLM API. The script must scan user-submitted prompts for obvious sensitive data patterns (SSNs, credit card numbers, specific project codenames) before they are sent to the model.

How to Execute

1. Use the `re` module to compile regex patterns for SSNs (e.g., `\b\d{3}-\d{2}-\d{4}\b`) and credit cards. 2. Create a function that takes a prompt string and returns a list of detected pattern violations. 3. Integrate this function as a middleware that either redacts or blocks the prompt. 4. Test with a suite of positive and negative test cases.

Intermediate

Project

Implement a DLP Gateway for an LLM Application

Scenario

Your company's internal chatbot uses a third-party LLM API. You must build a gateway that inspects both prompts and completions in real-time, applying different policies (e.g., block PII in prompts, redact proprietary terms in completions).

How to Execute

1. Use a framework like FastAPI to create an API proxy. 2. Integrate a more advanced DLP library (e.g., Microsoft Presidio) for named entity recognition on the prompt payload. 3. For completions, implement a post-processing hook that scans the streamed response and applies redaction rules. 4. Log all detections to a SIEM and generate alert thresholds.

Advanced

Project

Design a Zero-Trust DLP Architecture for Agentic AI

Scenario

You are the security architect for a company deploying autonomous AI agents that can read internal databases and execute code. A single prompt injection could lead to mass data exfiltration via the agent's completion stream.

How to Execute

1. Implement a 'prompt firewall' that deconstructs complex agent instructions into atomic sub-tasks, applying least-privilege data access checks at each step. 2. Use ML classifiers trained on your company's data to detect novel or obfuscated exfiltration attempts in completions (e.g., data encoded as stories, poems). 3. Integrate with a confidential computing environment to process sensitive prompts in a trusted execution environment (TEE). 4. Develop a real-time 'kill switch' protocol that can halt and rollback agent sessions upon anomaly detection.

Tools & Frameworks

Software & Platforms

Microsoft Presidio (NER-based detection)OpenDLP / Custom regex enginesAPI Gateway Policies (AWS API Gateway, Azure APIM)LLM Observability Platforms (LangSmith, Arize)

Presidio is used for context-aware PII detection. API gateways are the enforcement point for DLP policies on API traffic. Observability platforms allow monitoring and tracing of prompt/completion flows for forensic analysis.

Frameworks & Standards

NIST Privacy FrameworkISO/IEC 27001 Annex A (Information Classification)MITRE ATLAS (Adversarial Threat Landscape for AI Systems)Zero Trust Architecture (NIST SP 800-207)

NIST and ISO provide the foundational data governance and risk management structures. MITRE ATLAS offers a specific threat matrix for AI systems. Zero Trust is the overarching security model for designing DLP controls.

Interview Questions

Answer Strategy

The candidate must demonstrate a layered approach. Start with deterministic checks (regex for structured patterns), then apply probabilistic NLP models (like Presidio) for unstructured text. Crucially, discuss context-e.g., the same string in a prompt labeled 'example_ssns.txt' vs. 'my_ssn' requires different handling. Sample answer: 'I deploy a two-stage pipeline: first, deterministic regex matching for high-confidence patterns like credit cards, immediately blocking or redacting. Second, I run an NLP-based entity recognizer on the remaining text, configured with context keywords (e.g., 'employee:', 'ID:') to reduce false positives. For ambiguous cases, I route to a human review queue rather than auto-blocking, logging the context for policy refinement.'

Answer Strategy

This tests problem-solving and trade-off management. The core competency is policy refinement and managing security vs. usability. Sample answer: 'I would first analyze the blocked prompts to identify the specific false positive patterns. I would then create a context-aware exception list, not a blanket rule. For example, if prompts starting with 'Brainstorming session:' trigger fewer false positives for certain patterns, I could apply a lower sensitivity threshold for that prompt category. I'd implement a 'shadow mode' for new policies where they log detections without blocking, allowing us to tune the model before enforcement. Ultimately, I'd work with the business unit to define 'allowed fictional contexts' and encode them as positive context signals for the DLP model.'