Privacy requires control over information as it crosses agentic boundaries. An agent’s permission to read a document does not automatically authorize it to send that document to a model provider, retain it in memory, disclose it in an output, or reuse it for another purpose. section 5.6
Exposure can expand as information moves through retrieval, prompts, model interactions, tools, memory, logs, traces, generated content, and evaluation datasets. Repository access controls alone cannot govern all of these transitions. Restrictions must travel with the information into derived artifacts and runtime uses. section 5.6 section 5.1.2
Set explicit controls for each use and transition:
- Limit information use to its authorized purpose and matter; prevent context or memory from entering another matter. section 5.6 section 5.1.2
- Restrict which providers and tools may receive sensitive information, where processing occurs, and whether generated content may leave the matter boundary. section 5.1.2
- Apply privilege, confidentiality, disclosure, memory expiry, and retention rules throughout the information chain. section 5.6
- Apply policy during execution: restrict retrieval, block disclosure, or route a decision for approval. Record the applicable policy and interventions in the decision trace. section 5.1.2
Privileged email in a legal dispute
A privileged email may be appropriate for a supervising lawyer’s review but prohibited from an external draft, persistent memory, or retrieval in another matter. The decision is not merely whether the agent can read the email; it is which subsequent uses are authorized and which must be blocked. section 5.6

























