Skip to content

Data Loss Prevention for AI: Stop Staff Leaking Data Into Chatbots Without Banning Them

Data loss prevention for AI, tested first: map how data reaches chatbots, send fake records down each route and buy only the controls your gaps need.

IT & Networks · 7 Nov 2026 · 11 min read

Data Loss Prevention for AI: Stop Staff Leaking Data Into Chatbots Without Banning Them

A close-up of a laptop keyboard with keys glowing blue in the dark

By@ashtech52Lead Full-Stack Engineer

No single product stops staff leaking data into AI tools, because the data gets there six ways: pasted prompts, uploads, personal accounts, browser extensions, desktop apps and AI features inside other software. Data loss prevention for AI that works is a tested mix of an approved tool, identity controls and inspection at the browser or network.

So don't start with a product. Almost every guide on this topic is written by someone selling one. Before you buy, send fake records through each route your staff really use, and let the routes that leak decide what you buy. You can start today with a spreadsheet and a test account.

This is the method we'd use in a cybersecurity hardening review. It's a design to run yourself, not a report of results.

Updated 4 November 2026.

Data loss prevention for AI is partial: blocking, inspection and detection all have gaps

Three findings from vendor documentation shape everything below.

  • Blocking is partial. Microsoft's own guidance says blocking unsanctioned AI apps applies on managed network connections, and that users can still reach them on unmanaged devices or personal networks.
  • Inspection is partial. Microsoft warns that apps including ChatGPT and Meta AI may sometimes send content in encoded form or through dynamic endpoints, which can affect enforcement. Its inline checks also stop at 4 MB of text and 3 MB per file.
  • Detection is conditional. A detector can miss the data you care about for ordinary reasons, as the test below will show.

None of that makes these tools useless. It means coverage is something you measure, not something you read on a data sheet.

What IBM's 2026 report shows about shadow AI, and what it doesn't

Shadow AI is staff using AI tools without the knowledge or approval of IT or security. IBM's Cost of a Data Breach Report 2026, conducted by Ponemon Institute and sponsored and analysed by IBM, is the figure everyone quotes. It's based on breaches at 602 organisations between March 2025 and February 2026.

The 43% gets quoted four ways. IBM's landing page doesn't mention shadow AI, and we haven't read the full report.

  • IBM staff article: the share of organisations with security incidents tied to unapproved AI use went from 20% to 43%.
  • Baker Donelson, a law firm: "incidents involving shadow AI" rose to 43%.
  • SureCloud, a vendor: 43% of security incidents, citing page 45.
  • Cycode, another vendor: 43% of breached organisations.

Those aren't the same claim. The study only covers breached organisations, so "of organisations" and "of breached organisations" probably mean the same thing here, but "of incidents" would be a different measure. We use IBM's staff wording, and suggest you check the report's shadow AI section before repeating the number in a board paper. SureCloud cites page 45, and Accuro AI, another summary, cites pages 46 to 47.

The related figures, all reported from the same study:

  • Cost: $5.39 million for these incidents, against $4.63 million last year. The overall average is $4.99 million.
  • Consequences: 49% of those incidents resulted in data loss or compromise, and 21% in regulatory fines, according to the IBM article.

What it doesn't show:

  • It's a study of organisations that were breached, so 43% isn't the share of all companies with a shadow AI problem.
  • A link between shadow AI and higher cost doesn't prove shadow AI caused the extra cost.
  • Nothing in the summaries we read says which route the data took. That's the question you need answered, and only your own test can do it.

Vendor data points the same way on behaviour. Netskope, which sells this kind of control, reported in January that 47% of AI users in its data used personal AI apps. Its July report says the fall in personal use stopped around March 2026, with 56% on managed apps only, 14% using both and 30% on personal apps only. An approved tool helps, but doesn't end the problem.

The six routes data takes: paste, upload, personal accounts, extensions, API, AI inside SaaS

  1. Paste into a prompt. The classic case: a customer email, a contract clause, a stack trace with a token in it.
  2. File upload. A whole spreadsheet goes across in one action, and size limits on inspection start to matter.
  3. Personal accounts. The same chatbot, the same web address, a different login. OpenAI's policy says it doesn't train by default on business products such as ChatGPT Enterprise and the API, while individual services can be used for training unless the user opts out. So the account decides the terms, and a URL block can't tell them apart without extra controls.
  4. Browser extensions. These can read prompts no matter which account is used. In December 2025 The Register reported Koi Security's finding that four extensions with over 8 million installs collected AI chats. In its account of Urban VPN Proxy, Koi said collection was on by default with no off switch, and that extension carried a Chrome Web Store Featured badge.
  5. APIs, desktop apps and add-ins. A developer's script or an IDE plugin never touches a browser. Browser controls won't see it. See our vibe coding security checklist and MCP explainer for how these tools connect to your systems.
  6. AI inside SaaS. Your CRM or office suite adds an AI feature, and data reaches it with no paste at all. The control is the vendor's settings and your contract. Permissions matter too: Microsoft says Copilot can only use content the user is authorised to access, so oversharing becomes an AI problem.

A related risk runs the other way: a model can be manipulated by what it reads, which our prompt injection examples cover. For data you deliberately feed AI systems, see data masking for AI.

The fake-record test you can run today

You need a test account on each tool, written approval from whoever owns security, and about an afternoon for a first pass. Logs can take up to a day to settle, so check back. Use no real data.

1. List your real routes. Ask three teams what they use, and pull a report from your proxy, DNS or SaaS discovery tool if you have one. Include personal accounts. Pick the top five tools.

2. Make your test records. Use three kinds in one small text file and one spreadsheet.

  • A canary string. Something unique, such as CANARY-7Q4M-2291. Add it to your DLP as a custom pattern if you have one. It can't be anyone's real data, and it only matches what you planted.
  • A fake customer row. An invented name, address and account number, in your real format.
  • Format-valid identifiers, for the types you care about, each tested with and without a label such as "SSN" or "card number" beside it.

3. Run a control first. Send the same records through a channel you know is monitored, such as email. If nothing fires there, your detector is the problem, not the route. This step saves you from false conclusions, because detectors have rules:

  • Test cards may be ignored. Microsoft says its credit card detector doesn't flag reserved numbers used for testing. Stripe's 4242 4242 4242 4242 may therefore pass straight through and look like a gap. Don't test with real cards either.
  • Labels change the result. Microsoft's US SSN patterns all need a keyword such as "SSN" within 300 characters. A bare nine-digit column in a pasted export may not fire. That's a useful finding in itself.
  • Confidence levels matter. A card number with no keyword or expiry date nearby is low confidence, and a UK National Insurance number without one is medium. If your policy only acts on high confidence, those pass.
  • Invented numbers may not match. The US Social Security Administration never assigns areas 000, 666 or 900 to 999, so those are safe to use, but a detector may ignore them.

4. Send the records through every route. Paste them, upload them, and repeat for each device type, browser, network and account. Keep it small, and only ever use the fake data.

5. Record what happened, then check again later. Log entries can lag: Microsoft allows up to 30 minutes for network events to appear, and up to 24 hours after you first connect a network integration. Chrome rules need about 5 minutes after a change.

Route

Device and network

Account

Stopped?

Logged?

Who was told, and when

Paste, Chrome

Managed, office

Corporate

Paste, Chrome

Managed, home

Personal

Upload, any browser

Personal laptop

Personal

Coding assistant plugin

Managed, home

Personal key

AI feature in your CRM

Managed, office

Corporate

Score each line as stopped, logged only or invisible. Invisible is the one that matters. It means you'd have no evidence if this happened with real data.

Control map: which route each control covers

This is our assessment from vendor documentation, not a lab result. "Partly" means it depends on set-up, so your test overrules this table.

Control

Paste

Upload

Personal accounts

Extensions

API and apps

AI in SaaS

Written policy and training

Partly

Partly

Partly

Partly

Partly

Partly

Approved tool people like

Partly

Partly

Partly

No

Partly

No

SSO on approved tools

No

No

No

No

No

No

Tenant restrictions at proxy

No

No

Yes

No

Partly

No

Browser controls and DLP

Yes

Yes

Partly

No

No

Partly

Endpoint DLP

Yes

Yes

Partly

No

Partly

No

Network DLP or SASE

Partly

Partly

Partly

No

Partly

Partly

Extension allow-list

No

No

No

Yes

No

No

Redaction before sending

Yes

Partly

Partly

Partly

Yes

No

Vendor settings and contract

No

No

No

No

Partly

Yes

Some details worth knowing before you compare quotes:

  • SSO governs the approved account only. It doesn't stop someone logging in elsewhere. That's the job of tenant restrictions: Anthropic's option has your proxy inject an allowlist header and needs TLS inspection. OpenAI documents a similar workspace block for ChatGPT Enterprise.
  • Browser coverage varies. Edge for Business applies Microsoft's policies across all profiles on Intune-managed devices, for a named list of apps. The list we read included ChatGPT consumer, Gemini and DeepSeek, but not Claude. Its Endpoint DLP paste control covers Edge, Chrome, Firefox and Safari, with extensions needed for Chrome and Firefox on Windows.
  • Unmanaged devices are the hard case. Most of these controls assume a device you manage. If staff use personal laptops, your test will show it.
  • Warn before you block. Microsoft advises starting in simulation mode, and Chrome's rules offer block, warn and audit.

What you may already own before buying a new tool

Check these before you take a sales call. Licence rules change, so confirm with your reseller.

  • Microsoft 365. Defender for Cloud Apps can find AI apps in use, and Purview offers Endpoint DLP, Edge inline controls and sensitivity labels. Microsoft notes that other AI apps can't process content encrypted with its labels, which is strong protection for your most sensitive files. Network data security needs Purview E5 or equivalent plus an integration such as Entra Internet Access, or Microsoft 365 E7.
  • Google Workspace and Chrome. Chrome Enterprise Premium, an add-on, includes a template that blocks pasting on sites in its generative AI category, with gemini.google.com excepted, so edit it if you haven't approved Gemini.
  • An existing proxy or SASE service. It may already inject tenant headers or inspect AI traffic. Ask for a demonstration against your own test file.

Then match what you buy to what the test showed.

What your test showed

Start here

Detector missed the control run

Fix patterns and confidence levels first. A new tool may repeat the problem

Personal accounts work on managed devices

Tenant restrictions, then browser policy

Leaks only in some browsers

Standardise the managed browser, then extend DLP

Desktop apps, plugins or scripts are invisible

Network-level inspection, and approved keys for developers

Unknown extensions present

Extension allow-list

Only personal laptops leak

Conditional access to company data, not more DLP

AI features in SaaS you hadn't reviewed

Vendor settings and contract review

Why a ban tends to push usage out of sight

Cloudflare, which sells these controls, concedes a ban may reduce use, then calls enforcement nearly impossible: staff have personal devices, outside networks and VPNs. It offers no measurements, so treat that as a view, though Microsoft's point about personal networks backs up the mechanism.

The real cost is visibility. A blocked user on a personal phone leaves no log in your systems, and pasting a customer list there is exactly the invisible result your test is trying to find.

So block tools, not AI. Microsoft's model does this: sanction some apps, block others, and inspect what goes to the ones you allow. Give people an approved tool that's good enough for the job, and a short rule with examples from their real work. Then rerun the fake-record test every quarter, and whenever you add a tool.

Frequently asked questions

What is AI DLP?

AI DLP is data loss prevention applied to AI tools. It finds sensitive information going to chatbots, coding assistants and AI features, then blocks, warns about or logs it. It usually works in the browser, on the device or on the network, and coverage differs by route.

Can traditional DLP see ChatGPT prompts?

Sometimes. Varonis, a vendor, says traditional DLP was built for email, file transfers and endpoints. Our reading is that it may miss text typed into a web page unless a browser or paste control is turned on. Microsoft and Google both now offer those controls. Test it with fake records, because encoded traffic and detector rules can still cause misses.

How do I stop employees pasting sensitive data into AI tools?

Give them an approved tool, then add browser or endpoint rules that warn or block on sensitive patterns. Restrict logins to your corporate workspace where you can. Explain what mustn't be pasted, with real examples. Then send fake records through each route to confirm that what you set up works.

Does an enterprise AI licence solve shadow AI?

No. It improves the terms for that account, such as no training by default on OpenAI's business products. It doesn't stop staff using a personal account or another tool. Netskope, a vendor, reported in July 2026 that 30% of AI users in its data used only personal apps.

Should we block ChatGPT?

Block specific tools you've rejected, not AI as a whole. A blanket block is hard to enforce: Microsoft notes staff can still reach blocked apps on unmanaged devices and personal networks. If you block a tool, offer an approved alternative at the same time, or usage is likely to move somewhere you can't see it.

Got a line marked "invisible"? Book a free consultation, bring your completed test sheet, and we'll say which gaps are worth paying to close, and which aren't.

Sources

Back to all insights

Keep reading.

Have a project in mind?

Tell us what you're building, what's slowing your team down or what you'd like to automate. We'll come back with honest next steps and a clear estimate.

No obligation and no sales script.

Popular searches

Change theme