Minimise before you send.
Retrieve the passages the task needs rather than whole documents. The cheapest data protection is not moving the data.
MKC2 results ↗An AI system creates derivatives of your data at every layer: prompts, embeddings, caches, logs, training sets. Each one is a copy with its own access control.
01 / DEFINITION
The short answer first.AI data security is the protection of data as it moves through an AI system, including every derivative the system creates: prompt payloads, embeddings, retrieval indexes, caches, logs and training sets. Each derivative is a copy, and a control applied only to the original protects none of them.
Access control on a document store is the part everyone implements. The gap is that an AI deployment reproduces the same content in five or six other places with entirely different access models, and those copies are created by the pipeline rather than by a person, so they rarely appear in a data inventory.
WORKED EXAMPLE
A service company wants to build a model from maintenance reports. The repair description is useful, but the source record also contains customer addresses, access instructions and contact numbers. Dropping those columns from the final training table does not establish that they were absent from intermediate exports, debugging traces or evaluation notebooks.
Choose one synthetic customer record and follow it through ingestion, parsing, cleaning, labelling, train-test splitting and registration. At each step, record the storage location, identity with access, retention rule and downstream consumer. Apply the boundary to derived artefacts as well as raw files. A label spreadsheet or failed-job payload can contain the same information the final dataset was intended to exclude.
Give a reviewer with ordinary model-development access the cleaned training data and check what they can reconstruct. Separately test an operator investigating a pipeline failure. The operator may need a correlation identifier and error class without needing a full customer address. Match the information exposed to the task being performed.
Delete the synthetic source under a controlled test and verify what happens to temporary files, embeddings, cached extracts and backups under their actual retention schedules. Distinguish immediate removal from a scheduled lifecycle transition. If a copy cannot be selectively removed, record that limitation and identify the technical owner. This exercise produces a concrete data-flow record and exposes gaps before real maintenance reports enter the workflow.
Ask a second operator to locate every copy created during the exercise using only the inventory. Any unaccounted export becomes a concrete investigation item for the next pipeline release.
02 / THE COPIES
| Derivative | Created by | Control it needs |
|---|---|---|
| Prompt payload | Every request | Transport security; no third-party transit if material is confidential |
| Embeddings | Indexing | Same access control as the source; treat as sensitive, not as numbers |
| Retrieval index | Indexing | Per-document permission metadata, filtered at query time |
| Response cache | Optimisation | Keyed by identity, never shared across users or tenants |
| Prompt and output logs | Observability | The strictest access control in the system; explicit retention |
| Fine-tuning set | Training | Curated provenance; consideration of what the weights may reproduce |
| Model weights | Training | Treated as a data asset, because they encode training material |
Two rows deserve emphasis. Embeddings are frequently treated as anonymous vectors and stored with looser controls than the documents they came from; they are a lossy but informative representation of the source text and belong under the same access policy. Logs are the most common real-world exposure: a debugging pipeline that records full prompts and responses recreates the entire confidential corpus in a store built for engineers.
03 / THE INFERENCE BOUNDARY
Using a hosted model means every prompt — including the retrieved documents inside it — is transmitted to and processed by a third party. Contractual terms about retention and training are meaningful and are a different kind of assurance from the material never leaving. Where the data is confidential by default, that distinction usually decides the deployment.
Privilege AI's work sits on the private side: local model execution with explicit hardware, network and operational constraints, and MKC2 running entirely offline for exactly this reason — material reviewed in a legal setting is privileged, and the simplest defensible statement about it is that it does not leave the environment. See private AI inference and on-premise AI.
Where hosted inference is used, the controls that matter are concrete rather than contractual: what the retrieval layer is allowed to place in a prompt, whether logging is disabled at the provider, which region processes the request, and whether the same data can be re-identified from the responses that come back.
04 / PRACTICAL CONTROLS
Retrieve the passages the task needs rather than whole documents. The cheapest data protection is not moving the data.
Filter retrieval by identity at query time. Filtering after generation is too late — the content is already in the context.
Prompt and output logs are the highest-concentration copy of sensitive material in the system, and usually the most widely readable.
One further control is worth naming because it is easy and rarely done: deletion has to reach the derivatives. When a document is deleted or a client relationship ends, the source record is removed and the embeddings, caches, logs and any fine-tuning set that absorbed it often are not. A deletion procedure that does not enumerate derivatives does not actually delete anything.
05 / QUESTIONS
They are a lossy representation and research has repeatedly shown that meaningful information can be recovered from them, particularly for short texts. The safe operating assumption is that an embedding is sensitive in the same way its source is, and should carry the same access controls.
It is necessary for debugging and it creates the most concentrated copy of sensitive material in the system. Log with the strictest access control you have, redact where possible, set an explicit retention period, and include the log store in every access review.
It may satisfy a legal requirement. It does not change the technical fact that the material was transmitted and processed elsewhere, and for data that is confidential by default that distinction is often decisive on its own.
Enumerate the derivatives first: source record, index entries, embeddings, caches, logs and any training set. Deletion has to reach all of them. Material absorbed into model weights is the hardest case and is a reason to be careful about what enters a fine-tuning set.
That is the cross-tenant case, and it is prevented at the retrieval layer rather than in the prompt: identity-filtered queries, no shared caches, and separate indexes where the isolation requirement is strict. See RAG access control.
WORK WITH PRIVILEGE AI
Know where your data goes.
We build systems where it is clear what data the AI uses and where every copy ends up.
Keep confidential material inside.
We can run AI within an environment you control, so sensitive data does not leave.
A clear data boundary.
The limits on your data are stated explicitly rather than assumed.
Talk with Privilege AI
Tell us which data enters your AI workflow and which copies are hardest for your team to account for.
CONTACT PRIVILEGE AI
Tell us which data enters your AI workflow and which copies are hardest for your team to account for.
Privilege AI builds retrieval and inference architectures where the data boundary is explicit rather than assumed.