One in five firms hit by shadow AI security incident

DNSFilter has published research indicating that one in five organisations have experienced a security incident linked to an unapproved AI connection. The findings point to a gap between rising spending plans for AI governance and what companies can detect on their networks.

It found that 40.3% of organisations would direct their next security dollar to AI security and governance, while 43% discover that an AI tool has access to their systems only after access has been granted, or do not discover it at all.

The figures come from a survey of 400 IT and cybersecurity professionals in North America at Manager level and above, working at companies with 150 to 5,000 employees. DNSFilter also compared the survey results with activity observed across its own network.

That comparison suggested a wider mismatch between perception and use. On a typical business day, the average organisation on DNSFilter's network connected to five distinct AI tools, while 15.2% connected to more than 10.

By contrast, only 0.5% of surveyed IT professionals said their employees used more than 10 AI tools, while 55.3% estimated the number at three to five. In one month alone, DNSFilter processed 3.4 billion requests to generative AI tools from more than 43,000 organisations across North America.

Approval gap

The research also examined how companies approve new AI tools seeking OAuth or API access to corporate systems. It found that 36.8% of organisations allow employees to grant access and notify IT afterwards, while 6.5% provide no notification at all.

That means IT teams at 43% of organisations learn about AI access only after it has been approved, or never. This creates a blind spot for teams responsible for security, procurement and software oversight.

Existing controls did not appear to close that gap. Although 54% of respondents said they use AI governance or monitoring controls to identify incidents involving third-party AI tools, 46.1% said an employee had still connected an unapproved AI tool to Google Workspace, Microsoft 365 or a single sign-on environment.

Among all respondents, 20.1% said one of those unapproved connections had already caused a security incident. That underpins the report's conclusion that roughly one in five organisations have already been hit by a shadow AI incident.

Budget focus

The study found that security budgets are still rising. Nearly three-quarters of respondents, or 74.5%, said their cybersecurity budget was larger than 12 months earlier, and 18.8% reported an increase of more than 10%.

AI security and governance ranked first when respondents were asked where the next incremental dollar would go. Cloud security placed second at 25.5%, almost 15 percentage points lower.

DNSFilter linked that spending priority to concern about hidden costs and unmanaged access, but awareness differed sharply by seniority. While 54% of all respondents said they were concerned that their organisation was incurring material AI costs invisible to procurement, software asset management and security monitoring, the figure fell to 43.5% among executive leaders.

That compared with 58.8% among IT management, making executives the least concerned group in the survey. The finding suggests the people signing off on budgets may be less worried than the teams handling implementation and oversight.

Tool oversight

The report also pointed to broader issues in security operations beyond AI. It found that 61.5% of respondents could not say all their security tools were properly configured, while 61.8% could not say all were actively monitored.

More than one in five respondents said half or fewer of their security tools were configured correctly. DNSFilter also found that 42.8% of organisations using its service connected to more than five distinct AI tools during a standard business-day window.

Ken Carnesi, Chief Executive Officer at DNSFilter, said the main problem is a lack of visibility.

"Organisations are spending to govern AI, but you can't govern what you can't see. On a normal business day, roughly one in seven organizations on our network connects to more than 10 AI tools, yet fewer than one in 100 IT leaders believe that's happening in their environment. In too many companies, the decision to connect an AI tool to corporate data is being made by someone who doesn't own the risk. Closing that gap has to be the focus of the next round of spending," Carnesi said.

Kasey Best, Vice President of Threat Research at DNSFilter, said the company's network data did not match what many organisations believed was happening internally.

"Unfortunately, what organizations self-report and what we see on those networks are often two very different numbers. Allowing AI is a legitimate policy choice, but making no choice at all is not. Every AI tool with web access is reaching out to domains for one reason or another, which means it's touching DNS, and that's where we can help. DNSFilter gives security teams a place to start answering three important questions: what's running, who approved it, and what happens the next time a tool asks for access to sensitive corporate systems," Best said.