How this editorial can be challenged
What should an institution delegate when it asks an AI assistant to act?
A personal account often bundles access accumulated across many roles; attaching an assistant to that identity can make a narrow request inherit a much wider ability to act. Task-scoped authority limits consequences even when intent or interpretation goes wrong.
Agents need context and useful permissions; repeatedly asking humans to approve tiny actions can erase the efficiency gain and encourage habitual clicking.
Bounded delegation is not approval of every keystroke. It means defining the allowed task, resources, duration and escalation boundary in advance, then verifying both performance and containment.
The device-code case is a controlled demonstration with prerequisites. GAO used older systems and did not estimate current failure rates. The transparency initiative has not yet demonstrated operational effectiveness.
Independent workflow tests showing that broad-identity agents deliver materially greater value while reliably containing unintended actions would weaken the case for tighter task boundaries. Repeated failures to keep actions within a granted scope would strengthen it.
The ordinary account becomes an institutional decision
Consider an employee asked to use an AI assistant on a small, legitimate assignment. The employee's account may already carry access to repositories, old projects and sessions that have nothing to do with today's job. Giving the assistant that identity can be convenient. It can also turn a narrow delegation into a much broader one without anybody deliberately deciding that the extra authority is necessary.
This is an institutional design problem before it is a science-fiction problem. The useful question is not whether the assistant wants power. It is whether an ordinary request arrives with permissions larger than the work requires. An organization can create that mismatch even when its staff, software supplier and model are all trying to be helpful.
A real login does not settle the real intention
Microsoft's device-code demonstration illustrates the distinction between a valid authorization and an intended one. In its controlled setting, a user approved an attacker-initiated request, with possible consequences shaped by enabled features and existing permissions. This was not a claim that a passive website visit compromises every coding agent. The conditions are precisely what make the institutional lesson useful.
A successful sign-in establishes something about identity. It does not necessarily establish that the person understood which session they were authorizing or what it could reach. Treating authentication as the end of the decision lets a familiar screen carry more meaning than it can bear. The interface and the permission system have to preserve the intended delegation, not merely recognize the account.
Refusal and containment do different work
The newly public GAO testing belongs to a different technical category: it concerns information provided by chatbots, not account permissions. It nevertheless challenges the temptation to treat a reassuring response as a complete control. Its older-model results cannot establish current failure rates. They do show why the scope of an assurance matters.
Our argument does not depend on assuming that every safeguard fails. Refusal can prevent some harmful assistance; authorization controls can limit some actions; external records can help investigate what happened. These protections answer different questions. An institution weakens its position when it uses evidence about one of them to make an unsupported claim about all the others.
Autonomy need not mean inheriting everything
The strongest objection is practical. An assistant that must ask permission repeatedly may save little time. Worse, a stream of routine confirmations can teach people to click without thinking. It would be perverse to answer a problem of meaningful authorization with a design that makes authorization meaningless through repetition.
Bounded delegation offers a different bargain. Define a useful unit of work and give the system room to perform it within specific resources and a limited duration. Crossing that boundary requires another decision. This is our proposed design principle, not a feature proven by today's releases. Its success should be measured by completed work, understandable exceptions and the consequences of mistakes—not by the number of confirmation boxes displayed.
The record should survive a confident explanation
Microsoft's transparency announcement points toward a helpful complement: evidence drawn from the runtime, identity and policy systems as well as the model's own reports. A human-readable explanation can aid diagnosis, but it should not be the sole account of an action. A customer needs to know what actually occurred and what authority allowed it.
The initiative remains a proposal for further work, not a demonstrated solution. We would judge the resulting systems by whether missing records, scope violations and contradictory accounts can be detected in realistic workflows. Evidence that broader permissions consistently deliver much greater value with reliably contained consequences would challenge our position. A roster of supporters alone would not.
The decision belongs before the first action
The organization buying an agent is also deciding how much of its own authority to package for automated use. That decision deserves to be explicit. Employees should not have to discover the real boundary by making an error, and customers should not have to infer it from a vendor's reassuring language. A narrowly useful assistant can be a substantial achievement; it does not become less impressive because its powers have an edge.
Our choice is to make the job, rather than the person, the unit of delegation. Let the system do enough to be useful, keep evidence of what it does and preserve a workable route to interruption. This will sometimes impose friction. The alternative is not frictionless intelligence; it can be friction deferred until an incident, when discovering what was authorized is slower, more expensive and much less fair to the people asked to carry the consequences.
Read the reporting
Opinion is ours. The factual record is linked below.
Microsoft — Coding-agent authorization demonstration Microsoft — AI Transparency Initiative GAO — Controlled testing and its limitations