---
title: "AI breaches don't start in the model; they start with who has access"
slug: estudio-brechas-ia-sin-control-acceso
lang: en
situation: gobernar
format: estudio
minutes: 5
updated: 2026-10-19
publisher: "IBM and Ponemon Institute"
published: "2026-07-29"
sources:
  - label: "IBM, press release: one in four malicious breaches are AI-enabled (Cost of a Data Breach 2026)"
    url: "https://newsroom.ibm.com/2026-07-29-ibm-study-one-in-four-malicious-breaches-are-ai-enabled,-costing-companies-6-million-on-average"
    date: "2026-07-29"
  - label: "IBM X-Force, analysis of the Cost of a Data Breach Report 2026"
    url: "https://www.ibm.com/think/x-force/2026-cost-of-a-data-breach-ai-adversaries-enterprise-risk"
    date: "2026-07-29"
---

# AI breaches don't start in the model; they start with who has access

> For anyone approving AI tools connected to company data who answers for the risk if something leaks.

## In 30 seconds

- IBM's annual data breach study finds that among organizations that suffered an AI-related breach, 92% lacked proper access controls for those systems.
- The AI risk in your operation is less exotic than it sounds: it is who gets in, with which permissions and through which connection.
- Before adding another tool, review access, connections and encryption on what already runs.

## The exhibit

```exhibit
title: Basic controls over AI are still the exception
unit: %
No proper AI access controls (among orgs with an AI breach) | 92
Use access controls on AI models and data | 40
Encrypt sensitive data at rest and in transit | 37
Apply agents to vulnerability management | 18
note: IBM and Ponemon Institute, Cost of a Data Breach Report 2026, July 29, 2026. 602 organizations breached between Mar. 2025 and Feb. 2026.
```

## Fact

- The study covers 602 organizations breached between March 2025 and February 2026. The global average cost reached a record $4.99 million, up 12% from the prior year.
- More than 20% of organizations reported a breach targeting AI models or applications. Main causes: compromised APIs, applications or plug-ins (27%) and cloud misconfigurations affecting AI workloads (27%).
- Of organizations with AI-related breaches, 92% lacked proper access controls. Only 40% reported using access controls on AI models and data.
- One in four malicious breaches was AI-enabled, at an average cost of $6 million.
- Only 37% encrypt sensitive data both at rest and in transit.

## My read

The AI risk conversation often drifts to sophisticated topics. This study points to something more basic: permissions, connections and configuration. That is good news for a manager, because it can be reviewed without being a model expert. A tool connected to email, the CRM or a shared folder inherits whatever permissions you gave it. If nobody knows what those are, that is the first risk worth closing.

## What to do on Monday

1. List the AI tools your area already uses, which data they connect to and who granted access. [Sheet: inventory of the AI you already pay for](/resources/inventario-ia-que-ya-pagas/)
2. Add three rows to your risk register: AI access, connections and plug-ins, and cloud configuration. [Sheet: risk and readiness register](/resources/registro-riesgos-y-readiness/)
3. Publish a short policy stating which data goes into no tool without approval. [Sheet: one-page AI policy](/resources/politica-ia-una-pagina/)

## What we don't know

- It only includes organizations that were breached; there is no comparison with those that weren't.
- IBM sponsors and publishes the study and sells security and identity products related to its findings.
- The release does not list countries; we don't know how many Latin American companies are in the sample.

## If this becomes a big decision

If the decision is hard to reverse and carries your name, the free tools at vicsol.org/decision-clarity/ help you structure it. They don't replace a formal review.

---

Vic Sol · Decision Desk · vicsol.org/resources/ · Free to use with attribution.
Independent summary of a public source; figures belong to their author. Not legal, financial or audit advice.