AI Breaches Often Involve Easy Targets, Not Sophisticated Model Attacks

The IBM Cost of a Data Breach report found that most AI-related breaches do not involve beating the model itself, but rather targeting vulnerabilities in its surrounding infrastructure.

A total of one in four malicious breaches were categorized as AI-enabled, with deepfake impersonation and AI-enabled malware being the leading types. However, only 40% of organizations had access controls on their models and data.

The most common ways attackers breached AI systems were through compromised APIs (27%), cloud misconfigurations (27%), and plug-ins or applications (25%).

The average cost of a breach was $6 million, with the United States setting a record at $11.5 million, nearly double the global average.