Where does our sensitive data go when agents send it to an LLM?

The app sees egress as bytes today. This page reads those flows as data type → agent → destination, so Security (also Compliance) can see which sensitive classes are leaving and where they land.

Lens

1 open PII leak

support-copilot sent customer records to an unapproved Gemini endpoint (generativelanguage.googleapis.com, not on the approved list).

Sensitive-Data Flow

allowedmonitoredflagged / blocked

Sensitive data

PIIhigh

customer records · 1,240

Secrets / API keyshigh

tokens, creds · 86

Source codemedium

repo snippets · 410

PHInone-detected

health data · 0

→

Agents

support-copilot

Support · Gemini

data-pipeline-agent

Data · Anthropic

claude-code-prod

Platform · Anthropic

cursor-ide

Eng · mixed

→

Destinations

Approved LLMsallowed

Anthropic, OpenAI

Unapproved LLM endpointflagged

generativelanguage.googleapis.com

Internal storesallowed

vector DB, S3 (in-VPC)

flagged pathPII (customer records)→support-copilot→unapproved Gemini endpoint142 records · 14m ago

Observed Flows

Each row is a data type leaving via an agent to a destination.

Data type → Agent → DestinationClassVolumeLast seen

PII (customer records) → support-copilot → unapproved Gemini endpointFLAGGED142 rec14m ago

Source code → cursor-ide → Anthropic (approved)allowed410 files2m ago

PII (customer records) → support-copilot → Anthropic (approved)allowed1.1k rec1m ago

Secrets / API keys → data-pipeline-agent → OpenAI (approved)blocked12 keys38m ago

Source code → claude-code-prod → Anthropic (approved)allowed210 files5m ago

PII (embeddings) → data-pipeline-agent → internal vector DBallowed8.4k rec3m ago

Sensitive Data Detected

Detections across all monitored agent egress this period.

PII

1,240

customer records

Secrets

tokens & API keys

Source code

410

repo snippets

PHI

none detected