AI Security Statistics 2026: What the Sources Show
Selected findings from IBM/Ponemon's 2025 breach study and OWASP's 2025 LLM risk list. IBM studied 600 organisations that had already experienced a data breach, so its percentages are not prevalence estimates for every enterprise.
Breaches involving AI models and applications
of the 600 already-breached organisations in IBM's study reported a breach of AI models or applications.
IBM/Ponemon, 2025of IBM's AI-model or application-breach subset reported that they had no AI access controls in place.
IBM/Ponemon, 2025Shadow AI and exposed data
of the 600 already-breached organisations reported a breach due to shadow AI.
IBM/Ponemon, 2025higher average breach cost among organisations with high shadow AI use than among those with low or no shadow AI use in IBM's study.
IBM/Ponemon, 2025of incidents involving shadow AI compromised personally identifiable information, versus 53% of breaches overall in IBM's study.
IBM/Ponemon, 2025of incidents involving shadow AI compromised intellectual property, versus 33% of breaches overall in IBM's study.
IBM/Ponemon, 2025Governance in the breach sample
of IBM's breached organisations had no AI governance policy or were still developing one.
IBM/Ponemon, 2025of organisations in IBM's study with an AI governance policy regularly audited unsanctioned AI use.
IBM/Ponemon, 2025Broader breach and LLM risk context
global average cost across all breach types in IBM's 2025 study. This is not an AI-specific breach average.
IBM/Ponemon, 2025OWASP places prompt injection first in its 2025 Top 10 for LLM applications. List position is not an incident rate.
OWASP LLM Top 10, 2025How to read these findings
IBM's 13% figure describes AI-model or application breaches within a sample of organisations that had already experienced a breach. The 97% access-control finding applies only to that AI-breach subset. Neither figure estimates AI-breach prevalence across all enterprises.
The US$670,000 shadow AI cost difference is an observed association, not a predicted saving from a policy or product. OWASP's LLM01 placement identifies an application risk priority, not the frequency of employee AI incidents or a control Aona claims to provide.
Methodology and primary sources
The 2025 Cost of a Data Breach study was conducted by Ponemon Institute, sponsored and analysed by IBM. It examined breaches at 600 organisations globally from March 2024 to February 2025. OWASP's 2025 Top 10 is a risk list, not an incident survey. This page reports selected publisher findings and no new Aona measurement.
Sources reviewed 26 September 2026.
FAQ
Questions about these statistics
How common were AI-model or application breaches in IBM's study?
What did IBM report about shadow AI breaches?
Is US$4.44 million the average cost of an AI breach?
Does OWASP say prompt injection is the most common AI incident?
How can Aona help a team assess employee AI exposure?
Test the AI use your team needs to secure
Bring one employee AI tool and action. See the observed activity, configured policy response and sensitive-data protection on a supported path, then agree what needs validation in your environment.