What Google Cloud DLP (Sensitive Data Protection) does
Google Cloud's Sensitive Data Protection, formerly known as Cloud DLP, is a data classification and de-identification service. It uses built-in and custom infoType detectors to inspect data at rest in services such as BigQuery, Cloud Storage, and Datastore, and it can also inspect content submitted directly to its API. De-identification techniques include masking, tokenization, format-preserving encryption, and redaction. The service's core methods, such as inspectContent and deidentifyContent, are invoked synchronously by an integrating application. It does not natively monitor live application, API, or agent tool-call traffic unless that traffic is explicitly routed through the API by the developer. This makes it well suited to finding sensitive information within structured and unstructured datasets, but it was not built as an in-line control for dynamic, interactive request patterns.
How MCP and agent tool calls differ from static content scanning
The Model Context Protocol (MCP) is an open standard, published by Anthropic, for how AI applications connect to external tools and data sources through a client-server architecture. It defines message exchange, based on JSON-RPC, for tool discovery, invocation, and resource access between an AI client and MCP servers at runtime. This is a session-based, dynamic interaction pattern, not a static data payload. Data classification tools were designed to detect and act on sensitive content patterns within a payload, not to authenticate an agent's identity or authorize which specific tool it may call in a given session. An agent might request comparable content across two calls while invoking entirely different tools with different access implications. Content inspection alone does not determine whether a given tool call should have been permitted in the first place.
Architecture comparison: two different control points
The table below summarizes the distinct layer each approach operates on within an AI agent stack.
| Control point | What it governs |
|---|---|
| Data content layer | Google Cloud DLP inspects and de-identifies sensitive content within data payloads and data stores. |
| Agent action layer | Trussed governs which tools an agent identity may call, and under what permissions, at runtime. |
| MCP interaction point | Model Context Protocol traffic is a dynamic, session-based pattern distinct from static content scanning. |
Where the two approaches complement each other
Data classification and runtime governance are not mutually exclusive controls, and a mature AI agent deployment will typically need both. If a tool call returns sensitive data, such as a customer record retrieved from a backend system, that data can still be classified and redacted before it moves further into an agent's context or is returned to a user, provided the response is routed through a service like Sensitive Data Protection. Runtime governance controls address a different question: whether the agent should have been permitted to call that tool at all, under which identity, and with what scope of permissions. Treating either control as a substitute for the other tends to leave a gap. Content-only inspection cannot enforce least privilege at the action level, and action-only enforcement does not itself classify or redact sensitive content that a permitted action still exposes.