AI Agent Least Privilege
A method for deriving the smallest workable capability set, and keeping it small.
MKC2 results ↗Some work cannot be done from outside the boundary. A privileged agent holds rights over systems that matter, which raises the standard for everything around it.
01 / DEFINITION
The short answer first.A privileged AI agent is an autonomous system holding rights over sensitive resources — a system of record, a production environment, confidential documents, or actions with external effect. The privilege is what makes it useful and what makes every other control load-bearing.
The distinction is not technical but consequential. An agent summarising public documentation and an agent with write access to a matter management system run identical software. One has a worst case measured in wasted time; the other has a worst case measured in disclosure or an action that cannot be withdrawn.
Most published agent guidance assumes the first case. The operating conditions for the second are different, and they are worth stating explicitly rather than inferring from general advice.
WORKED EXAMPLE
An infrastructure assistant can inspect deployment health and prepare a rollback. Restarting a production service is a more powerful operation than reading its status. If both actions use one long-lived administrator credential, the application loses the distinction between an engineer asking for information and an engineer authorising a change. Design the workflow around the privileged operation itself.
Start with a specific incident. A new release produces errors in one region, and the on-call engineer wants to restore the preceding version there. Identify the service, target region, exact release, approving identity and time window. Bind the temporary grant to that operation. Permission to restore one service in one region should not authorise unrelated restarts, changes to secrets or deployment of a different image.
Ask the agent to inspect another region under its ordinary read access. That may be allowed. Then ask it to restart that region using the first approval. The deployment service should reject the action because the target is different. Repeat with the correct target after the grant expires. A correct plan from an earlier conversation should still fail when its authority has expired.
Store the approved operation and the service's final response under the same correlation identifier. An investigator should be able to reconstruct what the person authorised, what the agent requested and what actually changed. Keep executable credentials out of model prompts and ordinary logs. An audit record that allows anyone reading it to replay the privilege defeats its own purpose.
Simulate the deployment service completing the change but losing the network response. The agent needs a way to inspect the operation status before retrying. Otherwise a timeout can produce repeated restarts or a rollback against the wrong release. Make the recovery instruction part of the approved procedure, and give the human operator a way to stop further attempts.
After the exercise, withdraw the elevated grant and test that normal health inspection still works. This establishes both sides of the design: the agent can accomplish its ordinary job after an incident, and its privileged capability disappears when the approved intervention ends. A successful rollback alone would not demonstrate either property.
02 / FIVE CONDITIONS
Privilege AI's position, drawn from building systems where the underlying material is confidential by default, is that an agent with elevated rights is operable when five conditions hold together. Any one missing changes the risk materially.
03 / WHERE PRIVILEGE IS UNAVOIDABLE
| Task | Why narrow access fails | Control that carries it |
|---|---|---|
| Cross-matter review | The question is about the whole corpus, not one file | Read-only, aggregate output, volume caps, full logging |
| Drafting from precedent | Requires access to prior work across clients | Identity-filtered retrieval; no cross-client leakage by construction |
| Reconciling systems of record | Needs write access to correct a mismatch | Proposal-only with human confirmation; reversible changes |
| Reviewing another agent's actions | Must see everything the first agent saw | Review-only authority; no ability to act on its own findings |
The last row is the case Privilege AI has built for directly. A supervision model that reviews AI-generated work and agent actions necessarily sees the same material as the systems it reviews, which is privilege by construction. MKC2 handles it with a hard ceiling rather than a narrower grant: it runs entirely offline, reports and flags, and does not act. Read access is broad; the authority attached to it is deliberately minimal. Details on the MKC2 page.
04 / THE FAILURE TO PLAN FOR
A compromised privileged agent does not trigger authentication failures. It uses legitimate credentials for legitimate operations in an illegitimate pattern.
This is the practical detection problem. Access logs show authorised reads by an authorised identity; the only anomaly is volume, sequence or timing. Detection therefore depends on the bounds and signals described in AI agent monitoring — records per run, scope-boundary events, egress attempts — rather than on authentication telemetry.
Containment depends on the same bounds. An agent whose credentials expire with the run, whose reach is capped, and whose outbound path is an allowlist has a bounded compromise. One with a standing service account and general network access does not, and the difference is set at design time rather than at incident time.
05 / QUESTIONS
Rarely, and never standing. Where an administrative operation is genuinely needed, the pattern that holds is a proposal the agent prepares and a human executes, or a just-in-time elevation scoped to one operation and one run.
A user's actions are deliberate and relatively few; an agent's are generated at machine speed from reasoning that untrusted content can influence. The same permission set is materially riskier in an agent's hands, which is why volume bounds and review-only defaults matter more.
It can, and it means every piece of context the agent reads reaches a third party. Where the material is confidential by default — privileged legal documents being the clearest case — that is usually the decisive argument for local inference. See private AI inference.
Every tool call with arguments, the identity used, a reference to the content that prompted it, the response, and the resulting action — enough to replay the run. Logging only the final output makes any later investigation impossible.
With the five conditions above, evidence for each, and a statement of the worst case and its reversibility. An argument based on model quality alone tends not to survive the question of what happens when the model is wrong.
WORK WITH PRIVILEGE AI
Agents with real rights, safely.
We build agents that can do real work while staying within explicit limits.
Decided together.
Permissions, access to information and private deployment are planned as one.
Built for confidential work.
Our approach comes from building systems where the material is confidential by default.
Talk with Privilege AI
Tell us which privileged operation your agent needs, who can approve it and how an operator could revoke it.
CONTACT PRIVILEGE AI
Tell us which privileged operation your agent needs, who can approve it and how an operator could revoke it.
Privilege AI builds systems where permissions, retrieval boundaries and private deployment are decided together.