Using AI at Work Without Weakening Your Cyber Security

AI is now part of ordinary office work: drafting emails, summarising documents, taking meeting notes. Each of those tools needs access to your information to be useful, and that’s what makes it a security question as well as a productivity one.

The government’s Cyber Security Breaches Survey 2025/2026 found that among UK businesses using, adopting or considering AI, only 24% had cyber security practices in place to manage the risks. Here’s how to be in that 24%.

Where AI creates risk

• Data leaving your control. The NCSC notes that public AI providers store what you type, and advises against including sensitive information.
• Shadow AI. Staff using free tools, browser extensions or note-takers on personal accounts, without anyone knowing.
• Oversharing made visible. Microsoft 365 Copilot can only show people what they can already open, but in most companies that’s more than anyone realises.
• Apps with too much access. Many AI add-ins ask for permission to read your mailbox or files, and one click from a user can grant it.
• Prompt injection. Hidden instructions in an email or web page that trick an AI tool. The NCSC calls it one of the most widely reported weaknesses in large language models.
• Better phishing. The NCSC expects AI to make it hard for everyone to tell whether an email or password reset request is genuine.

Seven steps to use AI safely

1. Give people a safe option

Shadow AI usually happens because the approved option is missing or hard to find. If you have Microsoft 365, Copilot Chat signed in with a work account is covered by Microsoft’s enterprise data protection, so make it the default and show people where it is. Then write down which other tools are approved. Our post on what goes in a written AI policy covers this.

2. Control which apps can reach your data

By default, Microsoft 365 users can let some apps into their own mailbox or files. Microsoft recommends allowing user consent only for apps from verified publishers, and you can turn on an admin consent workflow so staff can request anything else. The NCSC recommends admin approval for high-risk integrations, only the permissions they need, and regular checks that they’re still needed.

3. Fix permissions before switching on Copilot

Microsoft says Copilot only surfaces data each user has at least view permission for, and stresses getting those permissions right. Remove old “anyone with the link” shares, tighten broad groups and close SharePoint sites nobody owns before you pay for Copilot.

4. Label and protect sensitive information

If you have Business Premium, Microsoft Purview sensitivity labels can encrypt your most confidential documents. Microsoft says Copilot honours the usage rights those labels give each person, so a protected document stays protected.

5. Limit what AI agents can do

AI that only drafts text is low risk. AI that can read incoming email, browse the web and then act, such as sending messages or moving files, is a bigger target for prompt injection. Give agents the least access they need, and keep a person approving anything that leaves the company.

6. Prepare for AI-written phishing

Phishing was the most common attack in the breaches survey, experienced by 38% of UK businesses. AI makes it more convincing, so:

• Turn on multi-factor authentication for every account, and use passkeys where you can. The NCSC said in April 2026 that passkeys should be the first choice wherever they’re offered.
• Never change bank details or pay on the strength of an email, call or video alone. Check through a phone number you already have.
• Train staff with realistic examples. Our post on AI-powered phishing has more.

7. Check AI providers like any other supplier

Ask where data is stored and processed, whether it’s used for training, what security certifications the provider has, and what happens to your data if you leave. Our checklist for vetting third-party apps works for AI tools too.

AI on the defending side

Attackers aren’t the only ones using AI. Security tools such as endpoint protection and email filtering use machine learning to spot threats too, and Microsoft 365 Business Premium includes Defender for Business. What matters is that alerts are seen and acted on. Our post on Microsoft Security Copilot looks at where AI-assisted investigation fits for a small business.

Make it routine

None of this is a one-off. Review approved tools, app permissions and Copilot access every few months, and whenever someone joins or leaves. Pair it with clear everyday AI rules for staff.

Our Cyber Security service covers the technical side: Microsoft 365 set up to best practice, email protection, multi-factor authentication and round-the-clock monitoring, with experts acting on real threats.