A new report on an earlier event
In an account published on 23 October of its 21 October London event, Microsoft describes trust, skills, security and human control as conditions for wider AI use. The timing matters: this is a newly published company account of an event earlier in the week, not an independently verified product launch today. [1]
Intelligence and authority are different
A separate Microsoft security post argues that an agent should have its own identity and narrowly defined permissions, rather than inherit all of its sponsor’s access. The company describes controls that can adapt permissions as risk changes. These are the supplier’s stated architecture and capabilities; they are not an independent audit demonstrating that every customer deployment is secure. [2]
Why the distinction matters at work
Consider a system asked to prepare an expense summary. Reading approved receipts is one activity; paying a supplier or changing bank details is another. A useful security boundary would keep those actions distinct even if a document tells the system to do something else. The editorial point is that accuracy of an answer and authority to perform an action are separate questions. A fluent explanation cannot substitute for an enforceable permission.
The evidence organisations should seek
The practical evaluation is whether controls work in the actual environment: which resources are reachable, who approves consequential actions and what records remain when something goes wrong. Buyers can ask for demonstrations of blocked actions and recovery, alongside demonstrations of productivity. Microsoft’s message is relevant because it recognises the problem inside an organisation’s own systems. Whether a particular product solves that problem depends on configuration, integration and evidence from use, not the event presentation alone.
