IBM's 2026 Breach Report Puts a Dollar Figure on Ungoverned AI: $1 Million a Breach

IBM's 2026 Cost of a Data Breach report finds that AI added roughly $1 million to the average cost of a breach, the first year the report has isolated "AI added cost" as its own measured line item rather than folding it into general breach costs. The finding underneath that number is more useful than the number itself: 92% of AI-related breaches occurred at organizations that had no access controls on their AI models at all.

That's not a sophisticated attack pattern. It's the most basic governance failure in the category, an AI model or agent left with no meaningful restriction on who or what could reach it, and it accounts for the overwhelming majority of the incidents IBM measured. Compare that against the more headline-grabbing AI security stories this year, novel exploit classes, zero-click browser hijacks, supply-chain compromises through OAuth tokens, and the contrast is stark: most of the actual cost is coming from organizations that skipped the basics, not from being outmaneuvered by something exotic.

What this means practically

If your organization is evaluating AI tooling and weighing which security controls to prioritize first, this data says access control on the model or agent itself, who can query it, what it can act on, under what conditions, is the highest-leverage control to get right before worrying about more sophisticated defenses. A perfectly injection-resistant model with no access controls is still, per this data, the more common way organizations actually get breached. Basic governance, done consistently, is outperforming sophisticated defenses applied inconsistently, at least in what IBM's dataset actually shows this year.