Shadow AI Detection
Detect unauthorized AI and LLM usage across your environment by monitoring connections to OpenAI, Anthropic, HuggingFace, and other AI endpoints.
Shadow AI Detection identifies when applications or users connect to AI/LLM services that have not been explicitly approved by your organisation. This is critical for BFSI companies subject to RBI, DPDPA, and SOC 2 requirements around data handling.
How It Works
- Endpoint library -- Aurva maintains a curated list of known AI/LLM API endpoints (OpenAI, Anthropic, HuggingFace, Google Vertex AI, AWS Bedrock, Azure OpenAI, Cohere, Replicate, and others).
- Traffic observation -- The data-plane agent monitors outbound connections and matches destination domains against the endpoint library.
- Payload inspection -- When a match is found, Aurva inspects the request for sensitive data patterns (PII, credentials, proprietary code).
- Classification -- Each detected AI connection is tagged as Approved, Unapproved, or Under Review based on your governance policy.
Configuration
Managing the Endpoint Library
Navigate to Settings -> AI Security -> Endpoints to:
- View the built-in endpoint list
- Add custom endpoints (e.g. internal LLM proxies, self-hosted models)
- Mark endpoints as Approved or Blocked
Approved vs. Blocked
| Status | Behaviour |
|---|---|
| Approved | Connections are logged but do not trigger alerts |
| Blocked | Connections trigger a policy violation alert immediately |
| Under Review | Connections are flagged for manual triage |
Alerts
When an unapproved AI connection is detected, Aurva generates an alert containing:
- Source service and namespace
- Destination endpoint and resolved domain
- Whether sensitive data was present in the payload
- Timestamp and volume
Alerts are routed via your configured Alert Routes to Slack, Jira, email, or SIEM.
Dashboard View
The AI Security dashboard shows:
- Total AI connections (approved vs. unapproved) over time
- Top services by AI API call volume
- Sensitive data exposure events involving AI endpoints
- Newly discovered AI endpoints not yet categorised
Related Pages
- LLM Access Governance -- policy-based controls for approved AI usage
- AIOStack -- open-source AI observability framework
- External Activity Policies -- create rules for outbound AI traffic