Buying an AI agent means evaluating what it can reach

Microsoft’s October 1 Digital Defense Report summary describes activity observed by its security teams, not worldwide attack prevalence. It reports AI use in parts of existing attack workflows. Its analysis places agent security in the connected system: data, tools, identities, permissions and infrastructure alongside the model. The coverage includes attribution and revocation. These are Microsoft’s observations and analysis, not an independently demonstrated defense guarantee. Microsoft’s official summary.

My reading is that a buyer’s central question should be what a proposed agent can cause in the specific business context. Imagine two otherwise similar assistants. One summarizes a document supplied by a person; the other can find documents across a team and change shared records. Their answers could look equally polished while their responsibilities differ substantially. A choice between them therefore needs an account of permitted reach, not only a judgment about the quality of their replies.

Consider a purchasing assistant asked to prepare options for a renewal. Reading the relevant contract, recommending a supplier and committing the organization to new terms would be three separate responsibilities in this hypothetical assignment. A buyer should be able to discuss which responsibility the product supports without assuming that permission for the first also grants permission for the others. My recommendation is to evaluate the proposed delegation at that level of specificity.

Context can also change while a task is unfinished. Suppose the person who requested a comparison moves to a different role before the work is complete. What should happen to the assignment and the authority behind it? The useful question is not simply whether the agent remembers the goal, but whether the original permission still applies. A product decision should make that continuity understandable. Otherwise, an assignment might appear valid because it began validly, even after its business justification has changed.

Responsibility after delivery deserves equal attention. If an output incorporates several recommendations and one consequential change, the organization needs to distinguish what was proposed from what was actually done. My recommendation is to ask whether a reviewer can reconstruct that distinction and identify the human decision associated with it. A statement that activity is recorded becomes more useful when it explains the question the record is intended to answer.

These are purchasing questions, not claims that any particular system has failed or met the standard. They also keep a threat report in its proper role. Observations of unwanted activity can inform the questions a buyer asks, but they do not automatically show how a product will behave in a different setting. The evidence needed for the purchase should address the proposed work and its boundaries directly.

For an enterprise reader, I would use the report as a reason to request a concrete account of reach, continuity of permission and responsibility for completed actions. Those answers help turn a broad security discussion into an assessable product choice. They do not eliminate uncertainty, but they clarify what the organization is delegating and which judgments it still needs to make.