48% of breaches involved a third party
Up 60% from the previous year’s dataset. Covers all breaches, not only AI‑related ones.
Verizon, 2026 DBIRTwenty figures on shadow AI, vendor AI, open models, agents and regulation. Each one is dated, linked to its primary source and labeled when a vendor paid for the research.
AI supply chain statistics measure how much of an organization’s AI is built, bought or downloaded from outside, and how well it is tracked.
In 2026, public sources show a third party involved in 48% of breaches, third-party systems making up 39% of AI at UK financial firms, and 45% of employees regularly using AI on corporate devices.
Four figures frame the rest. Each is reported as its source published it, with the scope stated. The full list of twenty follows, grouped under five findings.
Up 60% from the previous year’s dataset. Covers all breaches, not only AI‑related ones.
Verizon, 2026 DBIRAmong AI systems reported by UK financial services firms, up from 33% in 2024.
Bank of England and FCA, 2026Regular users of AI on corporate devices, authorized or not, up from 15% a year earlier.
Verizon, 2026 DBIREnough to backdoor language models from 600 million to 13 billion parameters in one study.
Anthropic, UK AISI and Turing, 2025An inventory only helps if it matches what people use. These figures measure the gap between the two. For the definition and the four ways it gets in, see What Is Shadow AI?
| Figure | What it measures | Source |
|---|---|---|
| Stat 0145% | Employees who are regular users of AI on their corporate devices, authorized or not, up from 15% a year earlier. Measured in the report’s 2026 dataset. | Verizon Business, 2026 Data Breach Investigations Report, Executive SummaryPublished May 19, 2026 · Source 1 |
| Stat 0267% | AI users who reach AI services through non-corporate accounts on their corporate devices, a slight decrease from the year before. | Verizon Business, 2026 Data Breach Investigations Report, Executive SummaryPublished May 19, 2026 · Source 1 |
| Stat 031 in 5 | Organizations that reported a breach due to shadow AI. High levels of shadow AI added an average of $670,000 to breach costs. Based on 600 organizations breached between March 2024 and February 2025. | IBM, Cost of a Data Breach Report 2025, press releasePublished July 30, 2025 · Source 2Vendor-sponsored study |
| Stat 0463% | Breached organizations in the same study that had no AI governance policy or were still developing one. Of those with a policy, 34% ran regular audits for unsanctioned AI. | IBM, Cost of a Data Breach Report 2025, press releasePublished July 30, 2025 · Source 2Vendor-sponsored study |
Most organizations buy more AI than they build, and understand it through what the vendor tells them. The third-party AI risk management hub covers what to do about it.
| Figure | What it measures | Source |
|---|---|---|
| Stat 0548% | Breaches that involved a third party, up 60% from the previous year’s dataset. This covers all breaches in the report, not only AI‑related ones. | Verizon Business, 2026 Data Breach Investigations Report, Executive SummaryPublished May 19, 2026 · Source 1 |
| Stat 0639% | Share of all AI systems reported by UK financial services firms that are third-party implementations, up from 33% in 2024. Based on 196 survey responses. | Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services 2026Published October 26, 2026 · Source 3 |
| Stat 0753% | Share of high-materiality AI systems at those firms that are third-party implementations, up from 35% in 2024. | Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services 2026Published October 26, 2026 · Source 3 |
| Stat 0881% | Respondents that rely on vendor assurance and contractual arrangements to understand their high-materiality third-party AI systems. 24% rely on deep knowledge of those systems, against 77% for systems built in-house. | Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services 2026Published October 26, 2026 · Source 3 |
| Stat 0972% | Financial institutions that are only partially aware of which of their vendors use AI. A further 9% have not assessed it at all. The sample size is not published. | Ncontracts, 2026 State of Third-Party Risk Management SurveyPublished June 4, 2026 · Source 4Vendor survey |
| Stat 1057.9% | EU enterprises using AI that bought it as ready-to-use commercial software, the most common route. 21.8% developed their own. 2024 data, enterprises with 10 or more employees. | Eurostat, The use of artificial intelligence (AI) technologies in the European Union: Key results, 2026 editionPublished 2026 · Source 5 |
Models, training data and the code that AI writes are supply chain components too. An AI‑BOM is how you record which ones a system depends on.
| Figure | What it measures | Source |
|---|---|---|
| Stat 11250 | Malicious documents that were enough to backdoor language models ranging from 600 million to 13 billion parameters. The backdoor tested was narrow, and the authors say it is unclear whether the pattern holds for larger models. | Anthropic, UK AI Security Institute and The Alan Turing Institute, A small number of samples can poison LLMs of any sizePublished October 9, 2025 · Source 6 |
| Stat 1221.7% | Average share of software packages recommended by open-source code models that do not exist. For commercial models the figure is at least 5.2%. Measured across 16 models and 576,000 code samples. | Spracklen et al., We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs, USENIX Security 2025Published March 2, 2025 (v3) · Source 7 |
| Stat 13205,474 | Unique names of non-existent packages the models produced in the same study. The paper examines how such names open the door to package confusion attacks. | Spracklen et al., We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs, USENIX Security 2025Published March 2, 2025 (v3) · Source 7 |
An agent is a model plus tools and permissions, and each connection is a component someone else built. Our guide to agentic AI security covers the risks OWASP names.
| Figure | What it measures | Source |
|---|---|---|
| Stat 1480% | UK financial services respondents with AI agents deployed; a further 14% plan to deploy them within 12 months. Only 2% report agents with broad autonomy. | Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services 2026Published October 26, 2026 · Source 3 |
| Stat 1580%+ | Fortune 500 companies using active AI agents built with Microsoft’s low-code and no-code tools, measured over the last 28 days of November 2025. | Microsoft Security Blog, 80% of Fortune 500 use active AI AgentsPublished February 10, 2026 · Source 8Vendor telemetry |
| Stat 1627% | Share attributed to compromised APIs, applications or plug-ins among the causes of breaches targeting AI models or applications. Cloud misconfigurations affecting AI workloads were level at 27%. More than 20% of organizations reported such a breach. | IBM, Cost of a Data Breach Report 2026, press releasePublished July 29, 2026 · Source 9Vendor-sponsored study |
| Stat 1792% | Organizations reporting an AI-related breach that lacked proper AI access controls. Roughly one in five organizations in the study reported an AI-related breach. | IBM X-Force, AI-powered adversaries and the enterprise risk challengePublished July 29, 2026 · Source 10Vendor-sponsored study |
Boards disclose AI risk, and regulators have started to name the buyer’s duties. See AI compliance frameworks, compared for the dates and the three frameworks side by side.
| Figure | What it measures | Source |
|---|---|---|
| Stat 1883% | S&P 500 companies that disclosed AI as a risk in 2025, up from 12% in 2023. Disclosure data as of December 2025. | The Conference Board, In the S&P 500, Disclosure of AI Risks Surges from 12% to 83%Published April 22, 2026 · Source 11 |
| Stat 19EUR 15M | Maximum EU AI Act fine for breaching the obligations of deployers of high-risk AI: EUR 15 million or 3% of worldwide annual turnover, whichever is higher. Rules for Annex III high-risk systems apply from December 2, 2027. | Regulation (EU) 2024/1689 (AI Act), Articles 26, 99(4) and 113, as amended by Regulation (EU) 2026/1744Published July 27, 2026 (consolidated text) · Source 12 |
| Stat 2019 | Critical ICT third-party providers on the list designated by the European Supervisory Authorities under the Digital Operational Resilience Act, which has applied to EU financial entities since January 17, 2025. | European Supervisory Authorities (EBA, EIOPA, ESMA), List of designated critical ICT third-party providers under DORAPublished November 18, 2025 · Source 13 |
This page is a synthesis of public sources. Cranium AI ran no new survey and used no customer data. Every figure was re-checked against its primary source on November 17, 2026.
Each figure links to the organization that produced it: a regulator, a statistics office, a research paper or the publisher’s own release. Nothing is taken from a news summary.
We keep each source’s wording and scope. Where a figure covers one sector or one country, the row says so instead of implying a global number.
Surveys, telemetry and studies paid for by a vendor carry a visible label. They can still be useful, and you should know who funded them.
Analyst forecasts, figures we could not confirm at the source, and research from companies we are compared with on this site.
Thirteen sources, each opened on November 17, 2026. The right-hand column shows which statistics each one supports.
These statistics accompany a 15-page research report that adds the analysis: what the findings mean, a field playbook and a self-assessment.
Enter your work email and your copy downloads instantly.
By submitting, you agree to our Privacy Policy and to hearing from Cranium AI. Unsubscribe anytime.
Nothing happened? Download the PDF here.
Research Report
Built for the security, risk and governance leaders who own AI supply chain risk.
These figures describe a market, not our customers. Cranium AI works on the part a buyer can act on, through the five moves of the AI Trust Loop. It does not see a model a vendor runs in its own infrastructure until the vendor discloses it, and it does not replace a TPRM suite.
Build the inventory the shadow AI figures say is missing: an AI‑BOM from your code repositories, and a Shadow AI view from connected SIEM activity.
Guardian inspects traffic between your AI applications and the models they call, and shows the tool calls your agents make.
Map each AI system to the EU AI Act, NIST AI RMF and ISO 42001, with attestations backed by citations.
Arena tests the AI applications you build against adversarial attacks before they reach production.
Request an AI Card from each vendor and keep the returned cards in My Vendors. AI‑BOMs export as JSON, CycloneDX, SPDX and XLSX.
Straight answers, kept current. Last updated .
Ask us anything →Verizon’s 2026 Data Breach Investigations Report found that breaches with third-party involvement rose 60% from the previous year’s dataset and reached 48% of all breaches. The figure covers every breach in the dataset, not only those involving AI. The report analyzed more than 22,000 confirmed breaches.
In the Bank of England and Financial Conduct Authority’s 2026 survey of UK financial services firms, third-party implementations made up 39% of all reported AI systems and 53% of high-materiality systems. Across the EU, Eurostat reports that 57.9% of enterprises using AI in 2024 bought ready-to-use commercial software, the most common way of acquiring it.
Verizon’s 2026 report found that 45% of employees are regular users of AI on corporate devices, up from 15% a year earlier, and that 67% of AI users reach those services through non-corporate accounts. In IBM’s 2025 study, one in five organizations reported a breach due to shadow AI, and 63% of breached organizations had no finished AI governance policy.
Research published in October 2025 by Anthropic, the UK AI Security Institute and The Alan Turing Institute found that 250 malicious documents were enough to backdoor language models ranging from 600 million to 13 billion parameters. The backdoor tested was a narrow one, and the authors say it is unclear whether the pattern holds for larger models.
The Conference Board reported in April 2026 that the share of S&P 500 companies disclosing AI as a risk rose from 12% in 2023 to 83% in 2025, based on disclosure data as of December 2025. Disclosure is not the same as control: it tells investors the risk exists, not that it is managed.
From thirteen public sources: regulators, a statistics office, peer-reviewed research and publishers’ own releases, each opened on November 17, 2026. Cranium AI ran no survey and used no customer data. Research paid for by a vendor is labeled, and analyst forecasts and figures we could not confirm at the source are left out.
Yes. Cite the original source for any single figure, since the number belongs to the organization that published it, and cite this page when you refer to the compilation. A suggested citation appears in the methodology section, and every row links to the source so you can check the figure before you use it.
See how Cranium AI discovers, observes, governs, secures and proves every model, agent and vendor in your enterprise.
Cookie Policy · Privacy Policy
Global Privacy Control detected. Opt-out request honored: marketing tags stay off in this browser.