Nearly half of surveyed organizations in the United States and United Kingdom have experienced an incident in which sensitive information was exposed through an AI tool, according to a November 3 research report commissioned by cybersecurity company BlackFog.

Sapio Research conducted the online survey on behalf of BlackFog in September 2026. It covered 400 IT security decision-makers at companies with more than 500 employees, split equally between the U.S. and U.K.

The survey found that 47% of the respondents had encountered an incident of sensitive data being exposed through an AI tool at their organization. Among respondents whose organizations reported an incident, 77% said a company-approved AI tool was involved, suggesting that restricting unapproved services alone may not prevent sensitive information from entering systems without adequate controls.

The research also identified gaps in organizations’ ability to track information shared with AI tools. 

While 62% of respondents could identify the users or departments sending data to these services, only 48% could provide records of the types of information that had left their organization through them.

The survey found that 39% of organizations believed AI adoption was moving faster than their ability to govern it. Of those respondents, only 27% had comprehensive AI controls in place.

What’s worrying is that more than half, 56%, said autonomous AI agents acting on employees’ behalf fell outside their existing controls. In the technology and telecommunications sectors, that figure reached 70%.

Approved AI Tools Create New Data-Governance Challenges

The findings point to an AI governance problem that extends beyond employees using unauthorized AI services. Even approved tools can create exposure if employees enter sensitive information without restrictions or if the organization cannot track what data is being shared.

BlackFog found that 98% of respondents had an AI usage policy, but enforcement varied. Some 35% had written rules specifying what employees should not enter into AI prompts, without technical measures to enforce them, while only 56% said comprehensive technical controls supported their broader AI usage policies.

Controls are weakest around proprietary design, creative and marketing content. For that type of information, 32% of respondents said they could detect or monitor it but could not consistently prevent it from being shared, while a further 20% relied primarily on written policies or training.

What CDOs Should Prioritize

For chief data officers (CDOs), the findings highlight the need to connect AI governance with data classification, lineage, access management and monitoring. Organizations need visibility into the information flowing through approved AI tools, as well as the actions taken by agents operating with employee permissions.

This becomes more important as AI systems move from generating responses to carrying out tasks across enterprise applications. CDOs should work with security and technology teams to establish who can authorize an agent’s access, what information it can use, how its activity is recorded and how sensitive data is prevented from leaving the organization.

A governance approach that accounts for AI risks, responsibilities and controls should cover approved tools, unsanctioned services and autonomous agents acting on employees’ behalf.

Limits of the Survey

BlackFog is a cybersecurity vendor specializing in data-exfiltration protection and commissioned the research. The survey does not identify the organizations involved, the AI tools used or when the reported incidents occurred.

It also does not establish whether exposed information became publicly accessible or was accessed by an unauthorized third party.