Can DLP tools detect shadow AI usage?

Partially, and the gaps are structural rather than configuration problems. DLP inspects content in the channels it monitors, so it catches file uploads and monitored web traffic reasonably well. It misses browser-native paste into an already-approved domain, extension-mediated access, personal and unmanaged devices, direct API usage with a personal key, and anything happening in a tab where the destination is on your allowlist.

The strongest version of the vendor case is worth stating. Modern DLP paired with a CASB does detect a meaningful share of AI usage, particularly file exfiltration to known model provider domains, and inline enforcement can block a paste before it lands. Several vendors have shipped AI-specific classifiers and destination lists. This is real coverage and it is not nothing.

It is insufficient for three reasons. Destination-based detection fails when the AI feature lives inside Slack, Notion, or Salesforce. Content-based detection fails when the sensitive material is a question rather than a document, and questions carry more strategic detail than most files do. And neither approach tells you anything about what came back, which is where output risk, hallucinated content, and IP contamination sit.

The category error underneath the question: DLP is built to prevent data leaving. Shadow AI risk includes data leaving, data being retained for training, outputs of unknown provenance entering your systems, and agents taking actions on your behalf. Three of those four are outside the DLP model entirely.

Akto Atlas covers the discovery and behavioral half, identifying which AI security tools are in use at the browser and extension layer and what data classes flow into them, which is the visibility DLP destination lists cannot produce.

Keep the DLP. Do not report its output as your AI inventory.