Microsoft Puts a Guard Between Outside Emails and Copilot. A New Survey Says Most UK Small Firms Never Formally Check the AI They Build.

On 13 October Microsoft published a post on its Copilot blog about "indirect prompt injection": instructions hidden inside an ordinary email, document or message that an AI assistant reads as a command instead of as text to summarise.

Microsoft's own example is the everyday one. You ask Copilot to summarise your inbox, one routine-looking email carries hidden instructions, and those instructions "may then influence model output". It says this can happen entirely within existing permissions, so passwords, access rights and data loss tools all work as designed and still do not stop it.

The new layer sorts content by where it came from. Emails from outside the organisation can be classed by policy as untrusted. Untrusted content goes through a separate, restricted step that keeps the useful facts and drops instructions before the main model sees it, and it is tagged with its origin.

Microsoft says the first focus is "high-risk content, including external communications". To use it, admins must "enable the applicable Outlook sender setting used to identify untrusted content" and follow Microsoft's guidance for CVE-2026-55145, a known Outlook Copilot flaw first published on 14 July and rated medium (6.3 out of 10). Microsoft adds that availability "may vary by release status and tenant environment", and gives no date for every account.

Microsoft is also plain that no single fix stops every attempt. This is one more layer, not a cure.

Pax8, a US company that sells cloud software through IT providers, published its Q3 2026 SMB AI Pulse Report on 12 October. It surveyed 1,150 leaders of firms with 5 to 499 staff online, 350 of them in the UK.

In the UK, 54% of firms using AI have built at least one AI tool now in regular use, such as an automation or an agent (software that carries out a task step by step on its own). Another 25% are prototyping.

Only 23% of those UK firms have a formal, documented review of what they build. That is the lowest of the four countries surveyed.

24% found AI tools, workflows or agents running without approval in the last six months. A further 13% are not sure, the highest uncertainty of any country.

Of UK firms paying for AI, 39% had at least one bill or usage charge higher than expected. Only 12% have a dedicated AI budget line, and 19% cannot see their AI spending separately at all.

44% of UK decision-makers spend three or more hours a week reviewing, editing or fact-checking AI-generated work.

Two caveats. Pax8 sells to the IT providers who would fix this gap, so it has a reason to find one. And the answers are self-reported, with a UK margin of error of about 5 percentage points.