IBM: 92% of Companies Hit by AI Incidents Lacked Basic Access Controls

IBM studied corporate security incidents in AI systems and found a pattern that shifts priorities in AI security: 92% of the companies that experienced such an incident had no basic access controls on their AI tools. In most of the recorded cases, the model itself (its architecture, weights or algorithm) was not the root cause of the problem.

Where the vulnerability actually lies

The key figure in IBM's study is 92%: that share of companies with AI incidents did not restrict who could access their AI systems or with what permissions. Access control, permission audits and monitoring of requests to the model were missing exactly where they should have formed the first line of defense.

Attackers, as a rule, did not go after the model directly. They used simpler vectors: they compromised an account with excessive permissions and gained access to the AI interface, and from there the system itself handed over the information they wanted.

• 92% of companies that fell victim to AI incidents did not have adequate access controls on their AI systems
• The AI model was rarely the root cause of the security breach
• The main attack vectors: excessive permissions for users and service accounts
• The lack of monitoring for anomalous requests to the AI system made the incidents worse

Access control for an AI system is not just a login and password. In a corporate setting, it means granular permissions: which data and knowledge bases the model can reach, and which employees and contractors may send queries, and how many.

Why AI systems have become an attractive target

Corporate AI tools increasingly get broad access to critical business data: customer databases, financial documents, internal correspondence and strategic materials. That is what makes them a valuable target. The reason is not vulnerabilities in the model itself, but the data that can be reached through it with ordinary queries.

A typical corporate AI deployment connects the model to documents, CRM data, work chats and knowledge bases. If nothing limits what the system can "see" and return in response to queries, the scale of a potential leak depends only on how broad the compromised account's data access was.

"The model itself is rarely the problem," according to the IBM study,

which points to organizational and configuration factors as the main cause of most recorded AI incidents.

Why basic measures still go unimplemented

The rush to adopt AI in the enterprise has created a situation in which AI systems are connected to production data before the necessary security perimeter is built around them. Teams often treat the tools as "just another SaaS" rather than as infrastructure with access to critical corporate resources.

The principle of least privilege, regular access audits, monitoring for anomalous requests and separation of test and production environments are not new concepts. IT security has applied them to CRM, ERP and databases for decades. According to IBM, organizations systematically skip this step when deploying AI systems.

What this means

IBM's study moves the conversation about AI security from algorithmic risk to organizational maturity. Public debate about AI threats often focuses on vulnerabilities in the models themselves: prompt injection, hallucinations, unpredictable behavior. But the real picture of incidents, according to IBM, is different: nine out of ten companies that suffered an AI breach could have avoided it with measures that information security specialists have applied for decades. For CISOs and IT teams, this is the key takeaway: before scaling AI infrastructure, make sure the same access control standards apply to it as to any other critical corporate tool.