Approved AI tools leak sensitive data

Nearly half of security leaders say their organizations have already seen sensitive data escape through an AI tool, and in most of those cases the tool was one the company had explicitly approved. That is the headline finding from a BlackFog survey published on November 4, which paints a picture of governance policies existing largely on paper while data keeps flowing out through channels nobody is watching closely.

The research, reported by Infosecurity Magazine on 6 November 2026, puts the figure at 47% of security leaders who have experienced sensitive data exposure through an AI tool at their organization. Of those incidents, 77% involved a company-approved AI tool rather than an unsanctioned one.

BlackFog's researchers say the results show that many organizations lack visibility into the types of data being shared with AI tools, even those approved for enterprise use. In other words, the problem is not only shadow AI. It is also the AI systems organizations have blessed, and then stopped paying attention to.

Approved tools, unmonitored content

The distinction matters because much of the shadow AI narrative has focused on employees quietly using unapproved tools behind the security team's back. BlackFog's numbers suggest a different dynamic: the exposure is coming through the front door, through tools that passed some form of organizational review.

The survey found that 62% of respondents could identify users or departments sending data to AI tools. That sounds reassuring until it is set against a lower figure: less than half, 48%, said they could provide any record of the types of data that had left the organization through AI tools.

So a majority of organizations can name who is using AI. Fewer than half can say what those users are actually putting into it. That gap between knowing the user and knowing the data is where the exposure lives.

Policy on paper, enforcement uneven

Almost every organization in the survey, 98%, reported having an AI usage policy in place. The ability to enforce those policies varied significantly across respondents.

Just over half, 56%, support their broader AI usage policies with comprehensive technical controls. Around a third, 35%, said their organization has a written policy specifying what information employees must not enter into AI prompts, but that there is no technical enforcement behind it.

That is a policy that depends entirely on employee compliance. Nothing in the technical stack stops a worker from pasting a client list, a source code snippet, or an internal memo into a prompt. The written rule exists; the control does not.

Pace of adoption outruns governance

The survey also asked security leaders about the speed of AI adoption relative to their governance capacity. 39% admitted that AI adoption is moving faster than they can govern it.

Among those respondents who feel outpaced, just 27% have comprehensive AI controls in place. The group that feels most behind is also the group with the least technical enforcement, a combination that leaves the least protected organizations the most conscious of their own exposure.

Agents outside the guardrails

The BlackFog research highlighted a separate governance gap around AI agents operating in organizational environments. Over half of respondents, 56%, revealed that autonomous AI agents acting on behalf of employees fall outside their organization's existing controls.

Technology and telecoms were the industries where this issue was most prevalent. 70% of respondents in those sectors stated that autonomous AI agents were outside of organizational controls.

The gap is notable because agents operate differently from a chatbot a person types into. An agent can take actions on behalf of a user, and if it sits outside the controls the organization has built, those actions may not be subject to the same scrutiny as a human's.

Detecting shadow AI is not the same as stopping it

Only 54% of security leaders said they had the ability to block or restrict unsanctioned AI tools. Around a third, 32%, said they are able to detect shadow AI use but are unable to consistently stop it.

That last group is in an awkward position: they can see the problem, but they cannot close it. Detection without the ability to block or restrict leaves a security team watching usage it has already decided is risky.

What the CEO says is missing

The finding that most incidents involved approved tools leads directly to a question about control over content rather than control over tool selection. BlackFog's founder and CEO framed the issue in those terms.

“This shows why security teams cannot stop at deciding which AI tools employees are allowed to use. Organizations also need to control what information is shared with them,"

— Darren Williams, founder and CEO of BlackFog.

That comment points at the gap the survey data describes: approval of a tool is a decision about the vendor, not about the data. The 48% who can produce records of what data left through AI tools are the minority; the rest cannot say what was sent.

The numbers behind the gap

The survey's findings can be read as a set of paired figures, each one showing a policy or capability on one side and a shortfall on the other.

• 47% of security leaders have experienced sensitive data exposure through an AI tool at their organization.
• 77% of those incidents involved a company-approved AI tool.
• 62% of respondents could identify users or departments sending data to AI tools, but only 48% could provide any record of the types of data that had left the organization through AI tools.
• 98% of organizations reported having an AI usage policy in place, while 56% support their broader AI usage policies with comprehensive technical controls and 35% have a written policy with no technical enforcement.
• 39% of security leaders admitted AI adoption is moving faster than they can govern it; of those, just 27% have comprehensive AI controls in place.
• 56% said autonomous AI agents acting on behalf of employees fall outside their organization's existing controls, rising to 70% in technology and telecoms.
• 54% of security leaders said they had the ability to block or restrict unsanctioned AI tools, and 32% could detect shadow AI use but not consistently stop it.

The research was published by BlackFog on November 4. The Infosecurity Magazine report on the findings appeared on 6 November 2026.

Why this matters for security teams

The most uncomfortable number in the survey may be the 77%. If most AI-related data exposure is happening through tools the organization approved, then the approval process itself is not the control that security leaders may have assumed it was. This suggests that tool-level vetting, however diligent, does not answer the question of what data is being sent, and that the answer may need to live at the data layer rather than the vendor layer.

For businesses, the gap between 98% with a policy and 56% with comprehensive technical controls is a reminder that a written AI policy is a starting point, not an enforcement mechanism. The 35% who rely on written rules with no technical backing are depending on employee judgment every time someone opens a prompt box. Whether that is an acceptable risk is a decision each organization has to make with its eyes open, and the survey data suggests many have not yet made it deliberately.

For the industry, the finding that 56% of organizations have autonomous agents outside their existing controls, and 70% in technology and telecoms, points to a governance problem that is arriving faster than the frameworks built to handle it. Agents act; a policy that only governs what humans type into a chatbot may not reach what an agent does on a user's behalf. Organizations that have not yet mapped where agents sit relative to their controls may find, as the survey suggests many already have, that the answer is outside them.

The exposure described in the BlackFog research is not a single breach or a single vendor flaw. It is a pattern of approved tools, written policies, and uneven enforcement, and the 47% who have already seen data exposed are the ones who can measure the cost of that pattern.