BlackFog survey: 47% report sensitive data exposure via AI

By Marcus Bennett

BlackFog’s research found that 47% of security leaders had experienced sensitive data exposure through AI tools. Among organisations that had reported such an incident, 77% said a sanctioned, company-approved AI tool was involved.

The 47% finding moves enterprise AI security beyond a simple question of approval. BlackFog’s results indicate that an approved service can still receive sensitive information without effective controls on inputs or data flows. The stakes are practical: a company can restrict the tools employees may use while still lacking the ability to see or stop what those tools receive.

BlackFog commissioned a survey of 400 IT security decision-makers at companies with more than 500 employees. The sample split evenly between the UK and the US. The survey covered governance policies, technical controls, visibility into AI data transfers, autonomous AI agents (AI systems acting on employees’ behalf) and unsanctioned tools.

BlackFog’s survey separates three practical questions: whether a rule exists, whether security teams can see activity, and whether they can prevent risky disclosure. The figures show large differences in the answers.

BlackFog said the findings suggest governance and technical controls are struggling to keep pace with AI adoption. In the survey, 39% of respondents said AI adoption was moving faster than their organisation could govern. Only 27% of those who felt outpaced said they had comprehensive AI controls in place.

That combination creates a specific gap between adoption and accountability. Respondents said AI adoption was outrunning governance, but fewer than half of the outpaced group had comprehensive controls. BlackFog’s research frames the problem as an enforcement problem rather than a lack of formal policy alone.

What does approval change?

The 77% figure is the report’s clearest warning against treating company-approved tools as inherently safe. BlackFog found that 77% of organisations with a data incident named a sanctioned tool as involved. The control problem can sit inside an approved workflow because approval establishes permission without necessarily providing visibility or prevention.

The 77% result also creates a coverage requirement. Security teams need to monitor approved tools and unsanctioned services with the same attention and extend oversight to AI agents acting for employees. That prescription comes from Dr Darren Williams, Founder and Chief Executive Officer of BlackFog, whose warning appears below.

Can security teams see what leaves?

Visibility remains incomplete. While 62% of respondents said they could identify the users or departments sending data to AI tools, only 48% said they could provide any record of the types of data that had left the organisation through those tools.

The two percentages measure different control objectives. User and department identification tells security teams where to investigate; a record of data categories would show what has already left. BlackFog’s results show that visibility of activity does not automatically provide visibility of content.

This gap appears alongside broad policy adoption. Almost all respondents, 98%, said their organisation had an AI usage policy. Yet 35% said they had a written policy specifying what employees must not enter into AI prompts, but no technical enforcement (controls that block risky disclosure). The survey also recorded that 56% supported broader AI usage policies with comprehensive technical controls