IBM and Ponemon studied 602 organizations breached between March 2025 and February 2026 and found that attackers didn’t need better tools — they needed unlocked doors. Global breach costs hit USD4.99 million this year, up 12% to a record high, while AI-driven attacks rose 56% and tacked on roughly USD1 million per incident. Ask how many of the AI-breached organizations had access controls in place before the attack, and the number drops to 8%.

The costs compound from there. Model inversion attacks, which pull training data straight out of a live AI model, average USD6.07 million each, and shadow AI incidents more than doubled to 43% this year, averaging USD5.39 million apiece. For a CDO, that’s not somebody else’s job to fix. The report treats data access, encryption and machine identity as the layer every AI breach actually ran through, and only 46% of organizations extend basic identity controls to the non-human accounts running their AI pipelines.

Download the report to learn how to:

• Close the access-control gap first. 92% of organizations breached through their own AI had no controls on it. Yet, it’s the fastest, cheapest fix in the entire study.
• Treat shadow AI as a live threat, not a policy footnote. Unapproved-AI incidents more than doubled to 43% this year and cost USD5.39 million on average, above the report’s global mean.
• Extend governance to machines, not just people. Only 46% of organizations secure the API keys and service accounts running their AI pipelines. The rest are one leaked credential away from a breach nobody notices for months.
• Point automation at prevention, not just response. Just 18% use AI agents for vulnerability scanning, exactly where frontier models move fastest. Everyone else is playing defense after the damage is done.
• Budget for post-quantum now, ahead of “harvest now, decrypt later.” Only 26% have a project underway, and AI training data is exactly what that attack is built to catch.