Third-Party Risk13 min read

IBM's 2026 Breach Report Just Measured Shadow AI in 16 Countries. Not One Was Caribbean.

By Nicholas Dunkley·Aug 19, 2026

TLDR

IBM and the Ponemon Institute's Cost of a Data Breach Report 2026, published 29 July 2026 from interviews with 602 breached organisations across 16 countries and regions, found workers using unapproved AI tools featured in 43% of security incidents, more than double the roughly 20% share recorded a year earlier.
Nearly 70% of the breached organisations in the study had no policy to govern or even detect AI use, and only 38% required IT approval before an AI tool went into service, down from 45% in 2025. Regulatory fines followed in roughly one incident in five.
None of IBM's 16 sampled countries is a Caribbean nation, and no CARICOM regulator, central bank, or industry body has published a comparable count of unauthorised AI use inside a Caribbean organisation. The region has never measured its own exposure.
Gartner forecast on 19 November 2025 that more than 40% of enterprises will have a security or compliance incident tied to unauthorised shadow AI by 2030, and Vanta's third-party risk data puts current shadow AI presence at 70% of companies already, discovering roughly 140 unapproved tools per organisation within 90 days of connecting monitoring.
StarApple AI's own live deployments, an AI-supported financial aid tool for working students and an AI-driven Sports Lab built with the University of the Commonwealth Caribbean, show what sanctioned, governed AI looks like in practice. CAIRMC's position: a Caribbean board cannot manage an exposure it has never counted, and counting it is now overdue.

Green coconut palm trees against a bright sky on a Caribbean coastline, no people visible

Photo via Unsplash

IBM spent a year interviewing more than 600 breached organisations across 16 countries and found that unapproved AI tools now show up in 43% of the security incidents those organisations reported. Not one of the 16 countries in that sample is in the Caribbean, which says more about where breach research funding concentrates than about any oversight in IBM's methodology. That infrastructure has not reached the region yet.

Shadow AI is the use of AI tools inside an organisation that security and compliance teams have not approved, cannot see, and have not assessed for data handling, and IBM's 2026 Cost of a Data Breach Report found it inside 43% of breaches across a 16-country sample, more than double the share a year earlier. No CARICOM regulator, insurer, or industry body has published an equivalent figure for a single Caribbean jurisdiction, which means the region is currently managing an exposure it has never measured.

This article sets out what IBM found, what Gartner and Vanta's separate data add to it, why the absence of a Caribbean number matters more than a bad Caribbean number would, and what StarApple AI's own sanctioned AI deployments in Jamaica show about the alternative to shadow use.

What IBM Actually Measured

IBM's Cost of a Data Breach Report, produced with the Ponemon Institute on a consistent methodology for 19 years, drew this year's findings from 602 organisations across 16 countries and regions that experienced a breach between March 2025 and February 2026, published on 29 July 2026. The headline cost climbed again: a global average of $4.99 million per breach, and containment now takes 247 days on average from first compromise to full resolution.

Buried inside that top-line figure is the AI finding that matters most for a risk committee. Workers using AI tools their organisation had not approved featured in 43% of the security incidents IBM studied this year, up from roughly 20% the year before. Those shadow AI incidents caused data loss or compromise in about half of cases, operational disruption in 40%, and a regulatory fine in roughly one incident in five. The governance side of the same dataset explains why the number moved so fast: nearly 70% of the breached organisations had no policy in place to manage AI or detect unauthorised use of it, and only 38% required IT approval before an AI tool went into service at all, a figure that fell from 45% the year before even as adoption accelerated. Fewer than one in five organisations had their governance and security teams coordinating on the problem.

Read plainly, the report describes the default condition of a workforce that adopted generative AI faster than any employer built the controls to govern it, across 16 of the world's largest and most heavily regulated economies. That is not a niche failure.

The Gap Nobody Has Filled

IBM's 16-country sample runs through the United States, the Middle East, Benelux, and a run of other major economies. It does not include Jamaica, Trinidad and Tobago, Barbados, or any other CARICOM member state, and it was never designed to. IBM samples where breach volume and enterprise security spend are highest, and no Caribbean jurisdiction currently meets that bar for a global vendor's research budget.

The absence matters because it is easy to mistake for reassurance. Nobody has counted the region's shadow AI exposure, which is a different thing from that exposure being low. CAIRMC's own review of the public record found no CARICOM central bank, financial regulator, data protection authority, or business association that has published a survey, audit, or estimate of unauthorised AI tool use inside a Caribbean organisation, whether in banking, government, or the private sector broadly. SOCRadar's CARICOM Threat Landscape Report 2026 tracks phishing and fraud volumes by sector in detail. It does not ask whether the staff being phished are also pasting client data into a free AI tool on their own account, which the Vanta and IBM data both suggest is now closer to the norm than the exception in comparable markets.

The PwC 2026 Caribbean Corporate Governance Survey, which polled 154 directors across the Bahamas, Barbados, Grenada, Jamaica, Saint Lucia, Trinidad and Tobago, and Bermuda, found only 6% of directors believe their board spends enough time understanding AI's impact, and just 9% believe they receive sufficient information on AI-related risks. That survey did not ask about shadow AI specifically. Given what it did find about board information flow generally, there is little reason to assume the boards least informed about AI's impact are somehow the most informed about which AI tools their own staff are already using without sign-off.

The obvious objection from a smaller Caribbean firm is that none of this applies at its scale, that shadow AI is a large-enterprise problem because only large enterprises have the headcount and the data volume to attract it. The mechanism argues against that comfort even where the IBM sample itself cannot settle it directly, since its 602 organisations skew toward the large, heavily regulated end of the market. Shadow AI does not require a large enterprise. It requires a laptop, a login, and a free AI tool one search away, which a five-person accounting practice has exactly as much as a regional bank does. The bank is simply more likely to have a security team that has already noticed.

Two More Numbers That Should Worry a Risk Committee

IBM's finding is not an outlier. Gartner forecast on 19 November 2025 that more than 40% of enterprises worldwide will experience a security or compliance incident tied to unauthorised shadow AI by 2030, drawn from a survey in which 69% of cybersecurity leaders said they already suspect or have confirmed evidence that employees are using AI tools their organisation has not sanctioned. "CIOs should define clear enterprise-wide policies for AI tool usage, conduct regular audits for shadow AI activity, and incorporate GenAI risk evaluation into their SaaS assessment processes," said Arun Chandrasekaran, Distinguished VP Analyst at Gartner, when the forecast was published.

Vanta's third-party risk data, drawn from anonymised monitoring of thousands of businesses between February 2024 and April 2026, puts the current state ahead of Gartner's 2030 forecast rather than behind it: 70% of companies already carry shadow AI somewhere inside their environment, and the average organisation discovers roughly 140 unapproved tools with access to its systems within 90 days of connecting monitoring for the first time. None of these three data sets, IBM's, Gartner's, or Vanta's, was built with a Caribbean respondent pool. All three describe a condition of adoption that has clearly reached the region too. Caribbean staff use the same free-tier AI tools everyone else does, on the same laptops connected to the same client and citizen data, inside organisations that IBM's governance numbers suggest are no more likely than their global counterparts to have a policy covering any of it.

What Governed AI Looks Like

The counter-example is not hypothetical. StarApple AI, the Jamaica-founded company that built the region's first commercial AI deployments, runs its current pilots with the University of the Commonwealth Caribbean under exactly the conditions shadow AI skips: a named institutional partner, a defined scope, and a governance relationship that exists before the system goes live rather than being reconstructed after an incident. One pilot is an AI-supported financial aid tool built to help working students manage their applications and eligibility. Another is a Sports Lab that uses AI, drone footage, and analytics to forecast game outcomes and support athletic programmes. Neither tool arrived on a student's or a coach's laptop through a personal sign-up. Both exist inside a partnership both institutions can name, describe, and audit.

"A financial aid tool touches exactly the kind of data a shadow AI incident puts at risk, income information, enrolment status, eligibility for support a student may depend on to stay in school," said Adrian Dunkley, founder of StarApple AI and Chairman of CAIRMC. "Building it as a named, sanctioned system with UCC rather than letting individual staff improvise around it with whatever free tool they found is the entire difference between a governed deployment and a shadow one. The technology risk is close to identical either way. The accountability is not."

Naming an owner is not a guarantee against every failure mode a financial aid tool or a Sports Lab can produce; a governed system can still misclassify an application or mis-forecast a game, and CAIRMC's position is that a partnership agreement is a starting control, not a finished one, and still needs its own testing and review. What a named partnership does reliably fix is the question IBM's data says most breached organisations cannot currently answer: who approved this, and who is accountable when it goes wrong.

Attribute	Shadow AI use	Governed deployment
Who approved it	An individual employee or manager, informally	A named institutional owner, before deployment
Data handling	Unknown to security and compliance teams	Defined and assessed against a data protection standard
Incident accountability	Unclear; no owner named in advance	Assigned to the institutional partnership that commissioned it
Visibility to the board	None, until a breach or audit surfaces it	Reported as a standing item under an existing partnership
IBM's 2026 breach-rate association	Present in 43% of incidents studied	Not the pattern IBM's data describes

The underlying model risk in each column can be identical. The same large language model can sit behind a sanctioned tool and an unsanctioned one. What changes is whether anyone can name who approved it, what data it touches, and who answers for it when something goes wrong. IBM's numbers say that single distinction now shows up in almost half of all breaches its researchers studied.

A Practical Shadow AI Register for Caribbean Boards

CAIRMC's recommendation is a register boards can start this quarter, not another warning, and it does not require waiting for a regional survey that does not yet exist.

Ask for a tool count before the next AI briefing. A board cannot govern what IT has not inventoried. Request a list of every AI tool with access to organisational data, sanctioned or not, before the next scheduled AI update, and treat any answer of "we don't know" as the finding itself rather than an acceptable gap in the paperwork.

Set the approval threshold IBM's data shows most organisations lack. Only 38% of the breached organisations in IBM's study required IT approval before an AI tool went into use. Requiring it is a low-cost control against a risk IBM ties to nearly half of the incidents it studied.

Treat data type, not tool brand, as the trigger for review. Any AI use touching client financial information, health records, or personal data covered under a Caribbean data protection statute should require the same sign-off StarApple AI's UCC partnership applies to its financial aid tool: a named owner, a defined data boundary, and a review before launch rather than after an incident.

Report shadow AI exposure as a standing risk register line, not a one-off memo. Vanta's data shows organisations that start monitoring find roughly 140 unapproved tools inside 90 days, not zero. A Caribbean board that has never looked should expect a comparable number, not silence, once it does.

Frequently Asked Questions

What is shadow AI?

Shadow AI is the use of AI tools inside an organisation that its security and compliance functions have not approved, assessed, or can see. It typically means employees signing up for free or personal-tier AI tools and using them with organisational or client data, outside any procurement or governance process.

What did IBM's 2026 Cost of a Data Breach Report find about shadow AI?

Published 29 July 2026 from interviews with 602 breached organisations across 16 countries between March 2025 and February 2026, the report found workers using unapproved AI tools featured in 43% of security incidents, more than double the roughly 20% share a year earlier. Nearly 70% of the breached organisations had no AI governance policy, and only 38% required IT approval before an AI tool went into use.

Was any Caribbean country included in IBM's study?

No. IBM's 2026 sample covers 16 countries and regions, none of them in the Caribbean. No CARICOM regulator, central bank, or industry body has published a comparable measurement of unauthorised AI use inside a Caribbean organisation, so the region currently has no baseline figure of its own.

What did Gartner forecast about shadow AI?

On 19 November 2025, Gartner forecast that more than 40% of enterprises will experience a security or compliance incident tied to unauthorised shadow AI by 2030. The forecast drew on a survey in which 69% of cybersecurity leaders said they already suspect or have confirmed evidence of employees using AI tools their organisation has not sanctioned.

How is StarApple AI's own AI deployment different from shadow AI?

StarApple AI's live pilots with the University of the Commonwealth Caribbean, including an AI-supported financial aid tool for working students and an AI-driven Sports Lab, are built under a named institutional partnership with a defined data scope and an owner accountable before launch. Shadow AI, by contrast, has no named institutional owner, no defined data boundary, and no visibility to a board until an incident or audit surfaces it.

What should a Caribbean board do about shadow AI this quarter?

CAIRMC recommends four steps: request a full inventory of AI tools with access to organisational data, sanctioned or not; set an IT approval requirement before any new AI tool goes into use, a control only 38% of organisations in IBM's study had; require a named owner and defined data boundary for any AI use touching client, financial, or personal data; and add shadow AI exposure to the risk register as a standing item rather than a one-off memo.

Why does the absence of Caribbean data matter more than a bad number would?

No measurement is not the same thing as low exposure. Vanta's monitoring data suggests which one is more likely: organisations that start looking find an average of about 140 unapproved AI tools inside their environment within 90 days, not zero, the more realistic starting assumption for a Caribbean organisation that has never checked.