AI IT Security Update | November 12, 2026
Before an AI assistant moves from drafting suggestions to working inside your business systems, make its authority visible. Pick one workflow and answer three questions: what can it read, what can it change, and where must a person approve the next step?
For an Orlando or Kissimmee business owner, this might be an assistant organizing customer inquiries, summarizing vendor documents or preparing follow-up messages. Start with one defined job instead of connecting every available inbox and folder.
This week’s risk: outside content can contain instructions
An AI workflow may read emails, attachments, websites or other material created by people outside your business. That content can contain instructions designed to redirect the assistant. This is called indirect prompt injection.
Microsoft’s security guidance describes how instructions embedded in third-party content can lead an AI system toward unauthorized actions. It recommends layered defenses, limited privileges and user verification of risky actions.
The operating rule is straightforward: a document being reviewed should not get to decide what the assistant is allowed to do. A supplier’s attachment cannot authorize a new recipient, broader file access or an unapproved business commitment.
Write a short permission brief
Before enabling a connected workflow, record the following with the employee or IT provider responsible for it:
- Purpose: the specific task and intended output.
- Inputs: the approved accounts, folders and document types.
- Allowed actions: for example, summarize or prepare a draft.
- Approval points: actions that must wait for an authorized person.
- Owner: who reviews problems and can stop the workflow.
Confirm the actual product, account type and data-handling settings. Do not assume that every AI feature bearing a familiar brand name has identical terms, protections or access controls. Keep confidential information out until the business has approved that use.
Separate preparation from business decisions
For an initial customer-inquiry pilot, let the assistant organize sample requests and prepare proposed replies. Have a person check the source request, the recipient and any factual claims before sending.
Put stronger approval points around actions with consequences: changing payment details, making purchases, sharing private files, deleting records, changing permissions or committing the business to terms. The reviewer should see the proposed action and affected information, not only a reassuring summary saying the workflow is ready.
Ask the administrator how those boundaries are enforced. An instruction written in a prompt is not a substitute for limiting the permissions available to the tool. If the platform cannot provide an appropriate boundary, keep that step outside the automated workflow.
Use the smallest practical access scope
Check the connected account and each integration separately. A task that reads one approved folder should not automatically inherit unrestricted access to the company’s document store.
Where supported, use read-only permissions during the pilot and enable write actions only after review. Check whether the assistant acts with a user’s permissions or a separate application identity, because those setups can expose different information and capabilities.
For Microsoft 365 Copilot, Microsoft states that it uses the permissions available to the user. That makes existing access worth reviewing: information already shared too broadly may also be available through the AI experience.
Run a small test and plan the stop
Use synthetic documents or approved, non-sensitive examples. Include an ordinary request, an ambiguous request and a harmless test document containing an instruction outside the workflow’s purpose. Ask your IT provider to help evaluate whether the controls hold.
Check that the assistant stays within its assigned sources, waits at the approval point and gives the reviewer enough context to spot a mistake. Record failures and fix them before widening access.
Finally, document how to pause the workflow, revoke its connections and review recent activity. Confirm what logs are available and who may access them. Avoid copying sensitive inputs into an unrestricted troubleshooting document.
Your owner action this week
Choose one workflow and complete its permission brief with the person responsible for IT. Keep it in a limited pilot until the approval and stop procedures have been tested.
To discuss the systems behind your AI workflow, explore local managed IT support, contact CMIT Solutions West Orlando & Kissimmee or call (407) 610-6720.
Sources: Microsoft Learn, Defend against indirect prompt injection attacks and Data, Privacy, and Security for Microsoft 365 Copilot. Sources reviewed November 10, 2026. Product capabilities and administrative controls vary by configuration.
