Skip to content

Data labels ​

Tags that travel with data through a workflow or agent, recording where it came from and what it contains, so later checks can decide what that data may do.

Inside an agent, every piece of text looks the same. A prompt built from a public web page is a string, just like a prompt typed by your own team. Data labels keep the difference. A label usually says two things:

  • Provenance: where the data came from, such as your own system, a signed-in user, or outside content like a web page or a tool result.
  • Contents: what kinds of data it holds, such as email addresses or card numbers.

Labels propagate. When a step reads labelled data, its output carries the same labels, so a summary of a web page is still outside content. Security people call this taint tracking or information flow control. Research defences against prompt injectionPrompt injection: Input that makes a language model ignore or change the instructions its application gave it, so it does something the application did not intend. build on it: the CaMeL design "associates, to every value, some metadata", including where the data came from and who may read it, and applies security policies to a value and its dependencies before a tool runs.

Example ​

A public webhook receives a support ticket. A step summarises it, and an agent gets the summary in its prompt. The webhook body was labelled outside content, so the summary and the agent's context are too. A rule that says "write tools need approval while the context holds outside content" now applies, even though nothing in the agent's prompt looks suspicious.

Common confusion ​

Labels do not detect attacks. They record origin and contents, and policies decide what to do with them. They also differ from document classification, where a file is tagged "confidential" once and the tag stays with the file. Data labels follow every copy, summary and excerpt of the data through the run.

In L2M ​

Each run labels where every step's data came from (trusted, user or untrusted_external), and tool policyMCP tool policy: A set of rules, enforced outside the model, that decides whether a given agent or client may call a given MCP tool with given arguments: allow it, deny it, or hold it for a person to approve. rules can test those labels (label_present). Optionally, switching on Track where data came from in a content rules rule set adds a scan for data classes, which a label_present sensitive check counts as sensitive, and keeps the labels with stored node results, memory and artifacts, so a later run that tracks labels inherits them (data labels).

Further reading ​

This website (not the L2M product) uses Google Analytics, only after you accept cookies. Cookie settings