AI data leakage starts when retrieval ignores who is asking
The short answer. Most AI data leakage inside a company isn't an attacker exfiltrating files. It's an assistant answering an employee's question with a paragraph from a document that employee was never allowed to open: a salary review, a board deck, an acquisition memo, another customer's contract. Nothing was hacked. The retrieval step found the most relevant text and handed it to the model, and relevance was the only test it applied. Similarity is not entitlement. Finding the right document and deciding who may see it are two different decisions, and leakage happens when only the first one is made.
Why this kind of leak is easy to miss
- It looks like the system working. The answer is accurate and useful. The only thing wrong with it is who received it, and nothing in the response flags that.
- The index reads as one account. Most retrieval pipelines reach the vector or search store with a single application credential, and that credential can read far more than any one employee.
- Permissions don't travel by themselves. A document's sharing settings live in the source system. When it's copied, split into chunks and embedded, those settings stay behind unless something carries them along and keeps them current.
Six places AI data leakage happens at retrieval
- The service-account view. Results reflect what the pipeline's credential can see, not what the person asking can see. The most common case, and it happens by design.
- Chunks without owners. Permissions are tracked per document, but the index holds chunks. A chunk that lost its link to its document's permissions is readable by everyone.
- Permissions that drift. Someone leaves a team or a folder is unshared, and the copy used for retrieval still says yes. Whatever rules you use, know how quickly a change in the source reaches the place where retrieval is decided.
- Filtering the answer instead of the inputs. Trimming the model's output, or hiding citations, is too late. Once a chunk is in the model's context it can shape the answer.
- Side channels. A withheld document can still leak through its title, a citation, a result count or a score.
- Shared caches and memory. A cached answer or a shared conversation memory can carry one person's results into another person's session (the MCP server post covers the shared-server case).
What this is not
This is not the same problem as data loss prevention. DLP looks at content leaving the company and decides whether it may go. Retrieval leakage happens earlier and inside the company: the right content reaches the wrong colleague. You need both, and neither one does the other's job.
What to test before you trust a retrieval pipeline
| Test | What a good answer looks like |
|---|---|
| Two users, same question, different access | Different results. The less-privileged user gets nothing they couldn't open in the source system. |
| Unshare a document, then ask again | The answer no longer draws on it, within a delay you know and have written down. |
| Break the permission check on purpose | Nothing comes back. An error returns nothing, not everything. |
| Ask about a document you can't open | The person asking sees no title, citation or count that gives it away. |
| Two users, one cache | One user's results never appear in the other's session. |
| Ask afterwards what was returned and withheld | A record per retrieval: who it was for, what was allowed, what was denied. |
Where Gateco fits
Gateco decides, per document and per chunk, what comes back from the retrieval systems you already run. Every candidate the index returns is checked against the requesting user's permissions, and anything denied is withheld before it reaches the model. When Gateco can't determine that a user is allowed, it returns nothing rather than guessing: deny by default, fail closed. Every retrieval records the decision and its inputs, so "what did the assistant retrieve for this person, and what was it refused?" is a query, not an investigation.
The permissions are the ones your organization already defines: groups and attributes from your identity provider, policies you write against them, and, where you choose to activate it, a source system's sharing settings imported from Google Drive, SharePoint or Confluence. A revocation made in Gateco applies on the next request; changes from a source system or a polled identity provider apply at its next sync, and SCIM changes apply as soon as your identity provider pushes them. It works across the vector and search backends you already use, without moving your data into a new index.
What Gateco does not cover. Gateco doesn't inspect content leaving your company or block uploads to outside AI tools; that's DLP. It can't take back an answer already delivered; it decides the next retrieval. It isn't a classification engine, and it doesn't scan for sensitive data you haven't permissioned. And it can't fix permissions that are wrong at the source: if you import a file's sharing settings and the file is shared with everyone, that is what Gateco enforces.
Common questions
What is AI data leakage?
Sensitive information reaching someone it shouldn't through an AI system. Inside a company, the most common path is retrieval: an assistant answers from a document the person asking isn't allowed to open.
Isn't this what DLP is for?
DLP controls content leaving the organization. Retrieval leakage happens between colleagues, before anything leaves, so it has to be stopped where documents are selected for the model.
Can't we just add a metadata filter?
A filter is a good start if every chunk carries correct, current permissions and the filter can't be skipped. The hard parts are keeping those permissions in line with the sources, handling group membership and per-user exceptions, failing closed when something is missing, and keeping a record of each decision. Why metadata filters aren't enough goes deeper.
Where does this fit in the OWASP Top 10 for LLM Applications?
LLM02:2026, Sensitive Information Disclosure, on its retrieval side, and LLM09:2026, Vector and Embedding Weaknesses, which covers access control on embedding stores. Our OWASP post maps both.
Want to see the check run? Read how permission-aware retrieval works.
If you have an AI pilot waiting for security sign-off, start free.
Sources
- OWASP Top 10 for LLM Applications 2026: LLM02:2026 Sensitive Information Disclosure and LLM09:2026 Vector and Embedding Weaknesses, GenAI Security Project.
More in this series
- AI agent security review: the checklist for moving agents to productionOctober 16, 2026
- OWASP LLM Top 10 (2026): what LLM02 and LLM09 ask of your AI retrievalOctober 23, 2026
- Agentic AI security starts with what the agent can retrieveOctober 29, 2026
- MCP server security starts with what the agent can retrieveNovember 6, 2026
Related reading
Ready to secure your AI retrieval?
Start with the free tier: 1,000 retrievals/month, no credit card required.