Last updated: 20 October 2026, 18:31 BST.
Microsoft Project Perception is expanding its security approach for AI agents inside organisations, following a company announcement on 19 October 2026. The proposals focus on identifying agents, limiting their authority and changing permissions when risk changes, rather than assuming company-deployed software is automatically trustworthy.
What Microsoft Project Perception adds
In its technical announcement, Microsoft describes controls which can preserve reading while blocking writes, narrow access or stop execution. It also describes separate agent identities and permissions scoped to the task. Guardian Agents are intended to help investigate risk and coordinate protective action within human-defined policies.
The company’s accompanying corporate explanation frames the issue as an insider-risk problem: an agent may start with a legitimate task and later encounter malicious instructions or act outside its purpose. These are Microsoft’s descriptions of its approach, not independent proof of effectiveness.
Why Brighton organisations should care
A local business assessing automation needs a clearer question than whether an AI model is generally capable. Can a tool which reads customer records also alter them? Can a booking assistant issue refunds? Who can remove that authority if behaviour changes? Those examples illustrate decisions an organisation should make; they are not reports of incidents at Brighton businesses.
The distinction between a person and their agent matters. Giving an assistant every permission its operator holds can create unnecessary exposure. A useful purchasing conversation starts with a list of the actions needed for the actual task, and asks how the product prevents everything else.
What you should do
- List existing agents, their owners and the accounts they use.
- Separate the ability to read information from the authority to change it.
- Ask a supplier to demonstrate how an unsafe write is blocked and logged.
- Check which announced features are available in your licence and deployment today.
For wider coverage, see our Technology & Innovation desk. This report does not recommend buying a particular enterprise bundle.
Brighton Locals’ view
The principle is sensible: security controls should be enforceable outside the model being supervised. The harder test is operational. Buyers need to see working boundaries, useful records and a recovery process, rather than accept the presence of another supervising agent as sufficient assurance.
What happens next
Availability details, administrator documentation and independent testing will determine how much the announcement changes everyday practice. A product description establishes intent; a successful demonstration under realistic conditions supplies stronger evidence.
Archive Microsoft campus sign, Redmond, April 2005. Derrick Coetzee / Wikimedia Commons, public domain. Not a photograph of the new security tools.
Leave a Reply