Nothing irreversible.
No filing, no sending, no committing to a system of record without a named person confirming it. The asymmetry between a wrong flag and a wrong action is the whole argument.
MKC2 results ↗Legal AI security is mostly about isolation: between matters, between clients, and between a system that reads client documents and a system that can act on them.
01 / DEFINITION
The short answer first.Legal AI security is the protection of an AI deployment inside a legal practice: keeping client and matter material isolated, enforcing ethical walls in retrieval, treating client-supplied documents as untrusted input, and bounding what any system may do to a live matter.
The general AI security surface applies unchanged. What is specific to a practice is that the isolation requirements are formal rather than administrative, that a large share of incoming documents originate outside the firm, and that the consequence of a wrong action lands on a client rather than on an internal process.
WORKED EXAMPLE
A legal AI system can enforce folder permissions at upload and still disclose material later through semantic search, a cached summary or an exported report. To test the boundary, create two fictional matters with similar parties and overlapping terminology but clearly different distinctive facts. Assign a reviewer to only one matter and keep the other available to an authorised control user.
Ask both reviewers the same question about a distinctive fact. Examine which documents were candidates, which passages entered the model context and which citations appeared in the response. A correct-looking refusal is insufficient if restricted text entered the wrong context. The control user should still be able to retrieve the authorised material, so the test verifies both isolation and ordinary usefulness.
Try the same request through a saved conversation, a document preview and a work-product export. Then remove access from the control user while an export is queued. Check whether the executor consults the current matter permissions or relies on an earlier decision. An asynchronous export can become the weakest boundary if it runs under a broad application identity.
Give the security owner identifiers for the matter, document, user, policy revision and decision. Keep the protected passages in a review location with the appropriate access controls. A general security dashboard should not become a second repository of confidential material simply because it is convenient for debugging.
Repeat the test after a connector migration or a restore. The data may be recovered correctly while group memberships, cache scope or export restrictions are rebuilt differently. Preserve the expected allowed and denied outcomes with the release evidence so the legal team can verify that the wall survived the operational change.
02 / ISOLATION
| Boundary | Requirement | Enforcement point |
|---|---|---|
| Between clients | No cross-client retrieval or caching | Query-time filtering, partitioned indexes, identity-keyed caches |
| Between matters | Access limited to matters a person is engaged on | Matter-level metadata on every chunk, filtered before scoring |
| Ethical walls | Specific people excluded from specific matters | Authoritative source checked at query time; ideally separate indexes |
| Between firm and client material | Client-supplied documents treated as untrusted content | Trust level in provenance metadata; no path to privileged actions |
| Between systems | A model that reads cannot also act | Separate components with a validated schema between them |
Ethical walls are the hardest of these to implement correctly, because they are exceptions rather than rules and they take effect immediately. A permission model built on inclusion — you may see matters you are assigned to — handles them only if the exclusion propagates to the index within seconds. Where that cannot be guaranteed, partitioning is the safer construction. See RAG access control.
03 / CLIENT DOCUMENTS ARE UNTRUSTED INPUT
A practice receives documents from clients, counterparties, courts and third parties continuously, and any of them can contain text intended for a machine reader. This is not a hypothetical concern about adversarial clients; content in disclosure bundles, filings and correspondence passes through many hands, and hidden text in document metadata or invisible layers is fully visible to a model while being invisible to a reviewer.
The consequence is that document intake is an untrusted input channel, and the controls are architectural: the component that reads incoming material should hold no privileged capability, and anything that can act should consume validated structured output rather than free text. See AI agent prompt injection.
Provenance and trust level belong on every indexed passage for the same reason. A pipeline that indexes firm precedent alongside counterparty attachments in one undifferentiated pool has combined sources with entirely different trustworthiness, and downstream handling cannot distinguish them. See RAG security.
04 / AUTHORITY ON LIVE MATTERS
No filing, no sending, no committing to a system of record without a named person confirming it. The asymmetry between a wrong flag and a wrong action is the whole argument.
A system that reports and flags has a worst case of wasted review time. MKC2 operates this way by design.
Every action and every retrieval attributable to the individual who requested it, not to a shared service account.
The records worth keeping follow from that: what was asked, which passages were retrieved, which model version produced the output, what the system flagged, and what the reviewer decided. That set answers the questions an incident or a professional enquiry will ask, and it is also the disposition record that eventually measures the system on live work. See AI agent monitoring and human-in-the-loop AI.
05 / QUESTIONS
Filter retrieval at query time against the authoritative source of the wall, and prefer partitioned indexes where the separation is absolute. Post-filtering an answer is too late, and a nightly permission synchronisation enforces yesterday's walls.
They are untrusted input. Text intended for a machine reader can be present in metadata or invisible layers, fully visible to a model and not to a reviewer. The control is architectural: the component that reads such material should hold no privileged capability.
Not without a named person confirming the specific action. The cost asymmetry is the reason: a wrong flag costs review time, a wrong filing or disclosure may not be recoverable at all.
The request, the identity, the passages retrieved, the model version, the output and the reviewer's decision. Those logs contain privileged material, so they need the strictest access control in the environment and an explicit retention period.
It removes external transmission, which is frequently the decisive question, and leaves internal access control, retention and supervision exactly as they were. Whether obligations are satisfied is a judgement for the practice.
WORK WITH PRIVILEGE AI
Keep each matter separate.
We design legal AI so information from one matter does not leak into another.
Controlled access.
Permissions and access to documents are decided together with private hosting.
Built for confidential material.
Privileged documents can stay inside an environment you control.
Talk with Privilege AI
Describe the matter boundary, roles and document routes you need to test before a legal AI workflow is released.
CONTACT PRIVILEGE AI
Describe the matter boundary, roles and document routes you need to test before a legal AI workflow is released.
Privilege AI designs retrieval boundaries, permissions and private inference together, for material that is confidential by default.