DNSFilter finds one in five surveyed IT and security managers report unapproved AI incidents

An AI tool can gain access to corporate systems before IT hears about it—or without any notification at all. DNSFilter’s The Cost of Shadow AI report, released November 3, found that one in five surveyed IT and cybersecurity professionals said an unapproved AI connection had caused a security incident, despite widespread budget increases.

The security vendor surveyed 400 IT and cybersecurity professionals at manager level and above in June 2026. Respondents worked at North American companies with 150 to 5,000 employees. The findings capture their accounts of spending, access and incidents; a separate portion of the research examines activity measured on DNSFilter’s network. Those are different forms of evidence, collected at different times.

The oversight problem starts when an AI tool asks for permission to connect to corporate systems through OAuth or an API—mechanisms for authorizing access or connecting software. In 36.8% of surveyed organizations, employees can grant that access and notify IT afterward. Another 6.5% send no notification. In those settings, IT notification is not a prerequisite for making the connection.

Overall, 46.1% said an employee had connected an unapproved AI tool to Google Workspace, Microsoft 365 or their single sign-on environment, which provides a shared login across services. The issue is therefore not only which AI applications employees use, but which ones receive connections to workplace systems.

These connections are not all harmless in respondents’ accounts: 20.1% said one had caused a security incident. Meanwhile, 54% said they use AI governance or monitoring controls to identify incidents involving third-party AI tools. Having a detection capability and requiring approval before access are distinct controls; the survey describes gaps around the latter even where monitoring is common.

The findings arrive alongside substantial reported growth in security spending. Nearly three-quarters of respondents, 74.5%, said their cybersecurity budget was larger than a year earlier. For 18.8%, the increase exceeded 10%. Asked where their next incremental security dollar would go, respondents put AI security and governance well ahead of cloud security:

• AI security and governance: 40.3% named it as their next spending priority.
• Cloud security: 25.5% selected it, almost 15 percentage points behind AI security and governance.

Uncertainty also extends to AI spending itself. Some 54% worried that their organization had material AI costs invisible to procurement, software asset management and security monitoring teams. Concern was higher among IT management, at 58.8%, than among executive leaders, at 43.5%. That measures differing levels of concern across roles—not an accounting of how much untracked spending actually exists.

61.5%Configuration uncertainty

Could not say all security tools in their environment were properly configured.

61.8%Monitoring uncertainty

Could not say every security tool was actively monitored.

DNSFilter’s network measurements add another perspective on visibility. In August 2026, the company processed 3.4 billion requests to generative AI tools from more than 43,000 North American organizations. Help Net Security reported that the report’s comparison of how many AI vendors organizations contacted used measurements across two business days that month.

The typical organization connected to five distinct AI tools on a standard business day, DNSFilter said, while 15.2% connected to more than 10. In the survey, just 0.5% of IT professionals estimated that employees used more than 10 AI tools for work. Most respondents—55.3%—put the number at three to five.

That contrast is suggestive, but it is not a before-and-after test of the same companies. The survey asked professionals to estimate employee use in June; the network figures measured connections among organizations using DNSFilter in August. They cannot show, by themselves, how accurately each surveyed manager understood their own organization’s activity.

Only 0.75% of measured generative AI requests were blocked, according to Help Net Security. That low share does not itself establish a security failure: DNSFilter’s vice president of threat research, Kasey Best, acknowledged that allowing AI is a legitimate policy choice. The unresolved distinction is between access an organization has deliberately permitted and activity it has not recognized well enough to make that choice.