Use Case: AI Security
LLM data leakage
An HR analyst pasted a spreadsheet of employee salaries and SSNs into ChatGPT to ask for formatting help. ManySignal's LLM DLP detected the PII in the API payload and blocked the request.
The DLP gap that ChatGPT created
Traditional DLP was built for a world where sensitive data moved via email, USB drives, or web uploads to known cloud storage services. ChatGPT, Claude, Gemini, and Copilot created a new exfiltration path that most DLP solutions don't cover: an API call that looks identical to any other HTTPS request, containing sensitive organizational data that users are voluntarily and unknowingly sharing with an external AI provider.
The Samsung incident (2023), where engineers shared proprietary source code and internal meeting notes with ChatGPT, made the front pages. Less publicized are the thousands of similar incidents that happen daily in organizations without LLM-aware DLP: HR teams analyzing employee data, legal teams summarizing contract terms, sales teams sharing customer information for CRM note generation.
rule: llm_sensitive_data_submission
type: dlp + network_proxy
sources:
- endpoint_proxy_agent (llm_api_calls)
- manysignal_sdk_middleware
- zscaler_casb (llm_category)
trigger:
- condition: |
api_destination IN [openai.com, claude.ai, gemini.google.com, copilot.microsoft.com]
AND dlp_classifier(request.body) == true
AND request.body.sensitivity_level IN ["PII", "PHI", "Financial", "Confidential"]
enrichments:
- dlp_findings: specific_data_types, confidence_scores
- user_context: role, data_handling_training_status
- request_volume: bytes_sent, conversation_history_size
- block_action: configurable (block vs warn vs log)
mitre: T1048 (Exfiltration Over Alternative Protocol)MITRE coverage
LLM data leakage FAQ
What counts as LLM data leakage?
Any instance where sensitive organizational data — PII, financial information, source code, internal documents, customer data, trade secrets — is sent to an external LLM provider's API without authorization or awareness. This includes: employees pasting customer data into public ChatGPT, developers submitting proprietary code to Copilot for review, and AI integrations that pass unredacted database query results to external models.
Does this apply to internal LLM deployments?
The primary concern is external LLM providers where data leaves the organization's control. For internal deployments (self-hosted Llama, private Azure OpenAI instances), data doesn't leave the environment, so the leakage risk is different. However, ManySignal monitors which LLM APIs are being called to ensure employees aren't bypassing approved internal models by using external alternatives.
Can ManySignal analyze the content of LLM API calls without decrypting the traffic?
For TLS traffic to external LLM providers, ManySignal uses process-level monitoring on the endpoint to capture API call content before encryption (via EDR process inspection or the ManySignal SDK middleware). This provides content visibility without requiring network-layer TLS inspection.
What DLP classifiers are applied to LLM inputs?
The same classifiers used in standard DLP apply: PII patterns (SSN, credit card, IBAN, passport numbers), healthcare data patterns (ICD codes, medical record numbers), financial data patterns, and custom classifiers for organization-specific sensitive data. The classifier runs on the API request body before it's sent to the LLM provider.
Add DLP inspection to every LLM API call
Real-time sensitive data detection in ChatGPT, Gemini, and Copilot submissions — before the data leaves your environment.