Skip to content
News

RESEARCH · September 29, 2026 · 8 min read

Shadow AI Statistics: What the Research Actually Says

Shadow AI is the term for staff using AI tools their organisation has not approved. The numbers below are the ones worth knowing, each read from the report that produced it — and two widely-quoted figures that could not be.

What people actually paste

The behavioural figures come from browser telemetry rather than surveys, which is why they are higher than the self-reported versions. People under-report this, and not dishonestly: pasting a paragraph into a chat window does not feel like moving data.

Most employees paste data into AI prompts. Not a minority, and not only technical staff.77% — the share of employees who paste data into generative AI prompts. LayerX, Enterprise AI & SaaS Data Security Report

The large majority of that pasting happens through personal accounts rather than company ones — which is what makes it invisible rather than merely risky. A corporate account leaves a log.82% — the share of those pastes made from unmanaged, personal accounts — outside any enterprise oversight. LayerX, Enterprise AI & SaaS Data Security Report

It is habitual rather than occasional. The figure that matters in that line is the second one: three of those pastes a day carry sensitive content.14 a day — average pastes into non-corporate accounts per employee, of which at least 3 contain sensitive data. LayerX, Enterprise AI & SaaS Data Security Report

Taken together, this has made AI the largest single route by which data leaves an organisation — ahead of the file-sharing and SaaS channels that security teams have spent a decade instrumenting.the largest — AI has overtaken SaaS as the single biggest channel for data leaving an organisation. LayerX, Enterprise AI & SaaS Data Security Report

The tools being watched are not the ones being used.

What it costs when it goes wrong

The baseline first, because every other number here is a comparison against it.$4.99 million — the global average total cost of a data breach across all industries. IBM, Cost of a Data Breach Report 2026

A breach involving AI costs about a million dollars more than one that does not.$6 million — the average cost of a breach that involved AI, about $1m above the global average. IBM, Cost of a Data Breach Report 2026

And AI is now involved in a quarter of malicious breaches. The word malicious is load-bearing and is usually dropped when this figure is quoted: the denominator is deliberate attacks, not every breach.one in four — the share of *malicious* breaches that were AI-enabled — a 56% rise year on year. IBM, Cost of a Data Breach Report 2026

Separately, more than a fifth of organisations were breached through an AI model or application itself, most commonly through a compromised API or a cloud misconfiguration.more than 20% — the share of organisations that experienced a breach targeting AI models or applications. IBM, Cost of a Data Breach Report 2026

The gap underneath all of it

Among breached organisations, barely a third encrypt sensitive data both at rest and in transit. This is the least surprising and most useful figure in the set: the failures are rarely exotic.37% — the share of breached organisations that encrypt sensitive data both at rest and in transit. IBM, Cost of a Data Breach Report 2026

Two numbers that could not be verified

Two figures appear constantly in coverage of this subject and are worth flagging, because a statistics page that repeats them without checking is the reason they spread.

  • •“Shadow AI adds $670,000 to the average breach.” Attributed to IBM's 2026 report. It does not appear in IBM's own announcement of that report, and until it can be read in the report itself it is being cited here as unverified rather than repeated.
  • •“Shadow AI incidents rose from 20% to 43% of breached organisations.” Same attribution, same problem. Both may well be in the full report; neither is in the public release.

Neither is implausible. That is precisely why they travel — a number that sounds right gets repeated, and after the fourth blog it has a life of its own regardless of whether anybody read the source.

What the numbers do not say

They do not say that people are careless. Somebody pasting a customer record into a chat window at four in the afternoon is trying to finish a piece of work with the best tool available, and the approved alternative is usually worse or does not exist.

Which is why blocking has such a poor record here. The 82% figure is the evidence: staff moved to personal accounts rather than stopping. A policy that makes the work harder converts a visible problem into an invisible one.

What actually reduces the number

  • •An approved tool that is genuinely as good, because the alternative to a good tool is a personal account rather than abstinence.
  • •Removing the identifying detail before it is sent, so that what leaves carries no one in it — which makes the paste harmless rather than forbidden.
  • •Instrumenting the browser rather than the network, since the traffic is TLS to a legitimate domain and looks like every other tab.
  • •Asking staff what they use it for. The answers are usually mundane and point at the gap in the sanctioned tooling.

Try Secure AI free

Frontier AI chat — encrypted by default, with your identity stripped before any request reaches the provider.

Common questions

What is shadow AI?

Staff using AI tools their organisation has not approved or does not know about — usually a consumer chatbot on a personal account, used for ordinary work. It is the AI version of shadow IT, and it spreads for the same reason: the sanctioned option is worse or does not exist.

How many employees use AI without approval?

Browser telemetry from LayerX found 77% of employees paste data into generative AI prompts, and 82% of that pasting happens through unmanaged personal accounts rather than corporate ones. Survey figures run lower, because pasting a paragraph into a chat window does not feel to people like moving data.

Does blocking AI tools work?

Rarely, and the evidence is in the figures themselves: the overwhelming majority of pasting already happens on personal accounts, which a corporate block does not touch. Blocking tends to move the behaviour off managed devices rather than stop it, which costs the organisation its visibility as well.

What does an AI-related breach cost?

IBM's 2026 report puts the average breach at $4.99 million and a breach involving AI at about $6 million. It also found AI was involved in one in four malicious breaches, a 56% year-on-year rise — note that the denominator there is malicious breaches, not all breaches.

Related

More from News
Secure AI

Use any AI. Stay anonymous.

© 2026 Secure Artificial Intelligence, Inc.Private details never reach the AI.
Shadow AI Statistics: What the Research Actually Says · Secure AI