Contents 5 sections
A recent academic test of 25 commercial retrieval assistants found data-extraction attacks against all 25, and in one case attackers pulled 41 percent of a 77,000-word book with 100 queries. That is an authorization problem that starts before the model writes a sentence.
The practical rule is simple: an AI answer must never show a document the user could not open. That means checking current rights before source content enters retrieval, context, tools, or the final answer, and the strongest design is to preserve each source system's native authorization model and enforce it at query time rather than depend only on a separate permissions index.
This is why the issue survives so many rollouts. Permission reviews are periodic, retrieval is continuous, and enterprise content does not sit in one place. HR closes an account, a group changes, a file share keeps an old copy, an embedding lingers in a vector store, and the assistant still answers from stale access because nobody made authorization part of the retrieval path.
Oversharing is what users notice, but stale access is the source
The scale is not theoretical. According to the ShareGate report, 29 percent of surveyed organizations said AI tools had already surfaced sensitive data that should not have been accessible. According to the Panzura blog, 38 percent of organizations reported leaving former employees or guests with access they should have lost during the prior year.
Those numbers describe different failures, but they meet in the same place. If the identity system says one thing, the source repository says another, and the retrieval index says a third, the assistant will eventually expose the version that is easiest to reach. The people who pay are the employee who sees a contract draft from another division, the customer team that exposes a case note, and the compliance lead who has to explain why a search box knew too much.
The persistence of the problem is also organizational. Application teams build indexes without identity, records-management, or security owners. Testing checks latency and answer quality but not unauthorized-answer rates, stale-group membership, or indirect leakage through snippets and citations. Meanwhile, broad indexing looks good in a demo, and narrowing retrieval looks like lost recall, even when it is the only defensible design.
Native authorization belongs on the retrieval path
The cleanest reference design keeps the source system as the authority and asks it, or its policy service, at decision time. That means connectors bring over content together with stable document IDs, ACLs, group references, sensitivity labels, and deletion events, then the retrieval layer filters candidates before any chunk is placed into model context. A published prototype described exactly that pattern, filtering retrieved chunks against document ACLs before context construction while still answering from the permitted parts of a request.
This also explains why a vector similarity index is not enough. An embedding can make restricted content discoverable even when the original file is protected, because similarity is not authorization. If the assistant can surface a title, a snippet, an entity name, or a citation from a file the user cannot open, the leak has already happened, even if the final paragraph sounds careful.
In practice, the decision path should be boring and explicit. Resolve the user's identity and effective memberships. Retrieve candidate documents. Check each candidate against the authoritative policy or source ACL. Build context only from the allowed set. If the permission state is uncertain, stale, or contradictory, abstain and send the user to reauthorization or human review.
That is also the point at which governance stops being a slide and becomes an operating control. Audit logging should record the user, document, effective groups, policy version, decision, and timestamp. Policy enforcement points should sit at retrieval, tool use, export, and citation boundaries. If an agent is reading from a file store, it should carry the user's identity or a narrowly scoped delegated token, not a universal service account.
Security has to be measured where retrieval happens
The recent permission-aware RAG proposal from SNU Elsevier Pure points to another useful choice: query the provider-controlled IAM endpoint rather than flattening every organization's rules into one new policy layer. That keeps the boundary tied to the system that already owns the entitlement model. It can add latency, and it depends on source availability, but it is harder for a stale copy to become the security boundary by accident.
That trade-off is real. Exact authorization checks, fresh group resolution, and source-system calls cost time and complexity, especially across document stores, wikis, email, case systems, databases, and tool gateways. But the alternative is worse: a fast answer that is wrong in a way the user can sue over, report, or forward outside the company.
The control set should be wider than retrieval alone. Behavioral analytics can flag unusual query volume, first-time access to sensitive repositories, or repeated probing of restricted topics. Redaction can help with partially authorized records, but it has to happen before model exposure and must block reconstruction from surrounding context. And testing has to include cross-user, cross-tenant, stale-permission, prompt-injection, indirect-inference, citation, export, and data-extraction cases, not just answer accuracy.
For teams putting this into production, the order matters:
First, inventory the systems of record that actually own permissions, including directories, HR lifecycle feeds, content stores, databases, and policy services.
Second, make retrieval consult those authorities at query time, not only a copied permissions table.
Third, enforce the same rule at tool calls, exports, citations, and conversation logs.
Fourth, measure unauthorized retrieval rate, stale-permission window, authorized recall, abstention rate, and p95 authorization latency together.
Fifth, treat denied access as a valid outcome, because a system that refuses safely is better than one that answers broadly and leaks quietly.
The harder cases are the ones leaders tend to postpone. Content copies, snapshots, embeddings, and caches can outlive the source permission. Metadata can be less protected than the document itself. And if a system cannot confirm identity, policy, or permission freshness, it should say it cannot answer, not confirm that a restricted document exists.
- 01Identity and lifecycleResolve who the user is and whether their current account state is active and eligible for access.
- Identity provider
- Directory groups
- HR lifecycle feed
- 02Source authorization intakeCollect content together with the permissions and labels that exist in the source system.
- Document connector
- ACL ingestion
- Sensitivity labels
- 03Candidate retrievalFind likely relevant records while keeping authorization checks separate from similarity search.
- Lexical index
- Vector store
- Tenant partitioning
- 04Exact policy checkValidate each candidate against authoritative rights before any chunk or tool result is exposed.
- Policy decision point
- Source IAM endpoint
- Authorization cache
- 05Answer construction and filteringBuild the response only from permitted evidence and block indirect leakage through snippets or citations.
- Context builder
- Language model
- Citation filter
- 06Audit and reviewRecord the decision and route uncertain cases to human review or reauthorization.
- Audit log
- Review queue
- Alerting feed
- Enforce identity and access checks at retrieval, tool, export, and citation boundaries.
- Log policy version, user, effective groups, document ID, decision, and timestamp.
- Abstain when identity, policy, or permission freshness cannot be established.
- Measure unauthorized retrieval rate, stale-permission window, and p95 authorization latency.
| Capability | Azure | AWS | Google Cloud |
|---|---|---|---|
| Identity and access management | Microsoft Entra ID | AWS IAM Identity Center | Cloud Identity |
| Event streaming for ACL and lifecycle changes | Azure Event Hubs | Amazon Kinesis Data Streams | Pub/Sub |
| Object and document storage | Azure Blob Storage | Amazon S3 | Cloud Storage |
| Search and retrieval indexing | Azure AI Search | Amazon OpenSearch Service | Vertex AI Search |
| Audit logging | Azure Monitor Logs | AWS CloudTrail | Cloud Audit Logs |
Governed retrieval is the difference between search and exposure
That is what makes permission-aware retrieval a management decision, not a model feature. You are deciding whether enterprise search is allowed to behave like a confused intern with access to every cabinet, or whether it will act like a trained employee who checks the badge reader before opening the file. The companies that get this right will be able to use enterprise AI search without turning it into a disclosure channel.
If you want the broader operating model behind that choice, our enterprise AI search guide covers the rest of the stack that makes retrieval usable in production. The same discipline also belongs in the governing architecture, because the real boundary is not the model, it is the route from identity to answer.
QueryNow builds the foundation enterprise AI runs on, and the agents that run on it. Talk to our team about the workflow you want to start with at Talk to our team every engagement starts with one scoped workflow and acceptance criteria agreed in writing before the build starts.
Sources
QueryNow leads AI transformation for enterprises in the US and Europe. Founded in Germany in 2014, we build the foundation enterprise AI runs on, and the agents that run on it, for pharma, healthcare, manufacturing and financial services organizations on Azure and AWS.
The Monday digest
The three posts worth your time each week, one paragraph each. From the people building production AI.



