Most security leaders are struggling to protect their AI systems due to inadequate tools and expertise, according to the AI and Adversarial Testing Benchmark Report 2026 by Pentera.

The study, which surveyed 300 US CISOs and senior security executives, investigates how companies are securing their AI and identifies critical shortcomings related to skill gaps and use of security measures that aren’t suited for the AI era.

AI systems are rarely independent. They’re often integrated into existing business technology, like cloud platforms, identity tools, applications, and data streams. Because different teams own various parts, central oversight is lacking.

Consequently, 67% of CISOs said they have poor insight into how AI is employed within their organization. No respondents claimed complete awareness; instead, they admitted to knowing about or accepting some level of uncontrolled or unauthorized AI usage.

Without a clear understanding of where AI operates or what it accesses, security teams struggle to effectively assess risk. Fundamental questions, like which identities AI relies on, what data it can access, or its behavior when defenses fail, often go unanswered.

Although AI security is now frequently discussed at the board level, the research indicates that the main problems aren’t financial.

CISOs listed the following as their biggest challenges in securing AI infrastructure:

• Lack of in-house skills (50%)
• Poor visibility into AI applications (48%)
• Not enough security tools specifically designed for AI (36%)

Only 17% considered budget limitations a major concern. This implies that many businesses are willing to invest in AI security but lack the necessary expertise to evaluate AI-related risks effectively.

AI systems have behaviors that security teams are still learning to analyze, such as autonomous decision-making, indirect access approaches, and privileged system interactions. Without proper skills and testing, it’s hard to determine if current controls are working as intended.

Due to a lack of AI-specific best security practices, expertise, and tools, most companies are expanding existing security controls to cover AI infrastructure.

The research revealed that 75% of CISOs use existing security controls, like endpoint, application, cloud, or API security tools, to protect AI systems. Only 11% reported using security tools specifically for AI infrastructure.

This pattern mirrors past technological changes, where companies initially adapt existing defenses before more specialized security measures emerge. While this may offer some basic security, controls designed for traditional systems might not account for how AI alters access patterns and creates new attack vectors.

Overall, the findings show that the difficulties of AI security stem from basic gaps rather than insufficient awareness or desire.

As AI becomes essential to company infrastructure, the report suggests that organizations should prioritize building expertise and enhancing how they validate security procedures in AI environments.

To fully understand the findings, download the AI and Adversarial Testing Benchmark Report 2026 for a more in-depth discussion of the data and important conclusions.

Note: This article was authored by Ryan Dory, Director, Technical Advisors at Pentera.