IBM report: AI attacks go through APIs and cloud, not models

In the 2026 edition of the Cost of a Data Break, compromised APIs and cloud configuration errors each account for 27% of the breaches that have reached an AI system.

The 2026 edition of IBM's Cost of a Data Break report, produced by the Ponemon Institute with 600 organizations, is a popular idea: the attackers largely bypass the models themselves to target the infrastructure around them. The compromised APIs and poor cloud configurations each account for 27% of the breaches that resulted in an AI system, far ahead of attacks directly targeting the model.

The most worrying finding is governance: only 40% of affected organizations had access controls protecting their models and data. Undeclared use of AI tools — the shadow AI — doubled as an aggravating factor, reaching 43% of cases.

The average cost of a violation rises to $6 million, up 35 per cent year-on-year, with peaks of 11.5 million in the United States and 6.64 million in health. The average detection and containment time is 247 days, the first recovery after five years of continuous decline. IBM's recommendation to safety managers is explicit: first secure the " plumbing" around the model, then automate detection, a leverage associated with approximately $2 million in savings per incident.