Data use control for AI agents

Control what
reaches the model.

An AI model cannot enforce permissions it was never given. Caber checks retrieved content against the requesting user's permissions and source history before it reaches the model.

Prevent data leaks, withhold stale passages, and scope agent access to each user.

Caber Console demo trace of one agent retrieval: sentences restricted by the user's policy struck through in red, duplicated passages struck through in blue, and stale results flagged.

One passage.
Different rules.

Three failures show up when AI agents work on enterprise data:

  • Data leaks. Block access to one copy, and agents find another.
  • Stale answers. A stale passage reads the same as a current one.
  • Overbroad agent access. A shared service account gives users access beyond their own permissions.

A relevant passage can still be restricted or stale. Caber links matching fragments across sources to find the context that retrieval alone leaves behind.

How Caber works

Read the docs
  1. Link matching content

    Fingerprint overlapping fragments roughly a sentence long. Link known copies by content across files, API responses, and agent tool calls.

  2. Resolve source context

    Associate each occurrence with its owner, access controls, timestamps, and byte ranges. Order related documents into a version history (beta).

  3. Enforce before inference

    Run compiled Rego in process using source context, the requesting user, and the purpose of the request. Release, mask, replace, or withhold individual fragments.

  4. Trace the decision

    Inspect source evidence and the deciding policies in the Console. Link fragments delivered to the model to its output (beta).

Run a demo on your own stack.

Start with the Caber examples or email us for a technical briefing.