Somewhere in a company, an AI agent is currently running. It has write access to the collaboration suite, the email inbox, and perhaps the code repository. And it’s operating without anyone noticing. In early 2026, Cybersecurity Insiders and Netskope surveyed a total of 1,253 IT and security professionals worldwide. What they found is clear: The adoption of AI has far outpaced security governance in most companies.
Governance lags behind implementation by 66 percentage points
A year ago, most companies viewed AI governance as something that could be formalized later, once its use had stabilized. That moment arrived sooner than expected. Copilots, code tools, and content generators were put to productive use in various departments even before the security department had established a framework.
“The adoption of AI has outpaced security governance: 73% of the companies surveyed are already using AI tools, yet only 7% have governance systems in place that enforce security mechanisms and policies in real time. This results in a structural gap of 66 percentage points.” – AI Risk and Readiness Report 2026, p. 2
Today, 68 percent of companies describe their AI governance as reactive or still in development. This means they take action only when problems arise, not before. 39 percent have already experienced a near-miss involving unintentional data disclosure. 17 percent of these companies did not make any changes afterward.
Behind these figures lies a specific organizational problem. More than a third of respondents report a fragmented AI implementation: Multiple teams are using tools independently of one another, without a common framework or uniform guidelines. One business unit operates autonomous agents based on informal agreements, while another has not even documented which AI tools its employees are actually using. Forty-eight percent believe that these very gaps will trigger the next major AI-related security incident.
More Money, Less Trust
90 percent of companies have increased their AI security budgets this year, with nearly one-third increasing them by more than 25 percent. Nevertheless, 29 percent feel less secure than they did twelve months ago.
“Paradoxically, this gap in AI governance exists even though companies are investing more than ever in security. This year, 90% spent more on AI security than in the past. Yet 29% feel less secure than they did twelve months ago. Investment is increasing, but confidence is declining.” – AI Risk and Readiness Report 2026, p. 4
The reason lies not in the budget, but in the architecture. Traditional security tools were developed for known file formats, predictable data flows, and human interactions. AI operates on a different level: It alters content semantically, acts autonomously, and communicates programmatically, without human involvement. Putting more money into old structures therefore means, above all, spending more resources on something that is ineffective against AI-specific risks anyway. Thirty-four percent of respondents cite economic pressure as the biggest obstacle: AI is being adopted faster than security structures can keep up.
Security teams can’t protect what they can’t see
Only 6 percent of companies truly know what’s happening in their AI environment: complete visibility into prompts, uploads, API calls, and communication between agents. 45 percent can at least see the officially managed applications; everything beyond that remains a mystery. 14 percent have no insights at all.
“You can’t protect what you can’t see. Only 6% of companies report having complete transparency regarding AI usage across their entire environment. 94% make decisions about AI security without complete information; for most, this is the norm.” – AI Risk and Readiness Report 2026, p. 5
Added to this is a technical vulnerability that the study classifies as the most serious of its kind: 88 percent of companies cannot reliably distinguish whether an employee is using an authorized corporate account or a personal account without data governance. On the same platform, using the same tool, yet with completely different consequences for data security. Access controls, DLP policies, and audit trails all rely on precisely this distinction. Without it, all other measures lose their foundation.
When an agent takes action, no one can stop him
AI agents already have write access to core areas in many companies: 53 percent to collaboration tools, 40 percent to email programs, 25 percent to code repositories, and 13 percent to cloud infrastructure. What happens if one of these agents performs an erroneous or malicious action? In 91 percent of companies: nothing—at least not in time.
“Agents have extensive access to corporate systems that is virtually impossible to prevent. If an agent initiates a malicious action, only 9% of companies can intervene before it is completed. 24% can stop some, but not all, agent actions; 35% would only find out about them later through logs; and 32% have no insight into agent actions at all.” – AI Risk and Readiness Report 2026, p. 8
The study describes a scenario that is already occurring in this form: An analyst arrives at work on Monday morning and notices that unusual changes have been made to access permissions. He traces them back to an agent who had created the relevant account 72 hours earlier and has been writing to production systems ever since. Every action is logged. No alert was ever triggered because there was no detection rule for agent-initiated operations. Thirty-seven percent of companies experienced AI-related operational issues in the past twelve months. For 8 percent, these led to system outages or data corruption.
AI describes what DLP no longer detects
Data Loss Prevention is considered a standard tool for preventing data leaks. The fundamental problem: DLP looks for patterns, but AI changes them.
“If an employee takes a confidential project description and asks an AI to ‘summarize it in a professional email,’ the AI might replace ‘Project X’ with ‘our upcoming strategic initiative.’ To a regex filter, ‘strategic initiative’ appears completely harmless, even though the semantic value (the confidential information) remains identical.” – AI Risk and Readiness Report 2026, p. 6
46 percent of respondents say that their systems overlook precisely this type of policy violation. 27 percent are unsure whether their tools even work reliably in such cases. Only 8 percent can confirm that their security systems semantically detect AI-transformed content.
What that means
In hindsight, 38 percent of respondents wish that governance measures had been put in place before the widespread adoption of AI. Only 7 percent are satisfied with their current approach—the lowest confidence indicator in the entire study.
Companies that take AI security seriously do not treat it as an extension of existing IT security, but as a discipline in its own right: complete visibility into all AI activities, real-time enforcement rather than retroactive enforcement, data control at the semantic rather than the syntactic level, and zero-trust approaches that also include agents and API keys. Those who ignore these variables and instead rely on the next tool will see the same numbers in the next report.
About the study: AI Risk and Readiness Report 2026. Cybersecurity Insiders in collaboration with Netskope. Sample: 1,253 cybersecurity and IT professionals surveyed in early 2026.
Cover Image © Wallap Puraya
