Decoding the world of cybersecurity

Employee AI misuse rivals phishing in security incident survey

A survey of UK and US cybersecurity managers identifies unauthorised AI use and data sharing as leading contributors to serious incidents or near-misses. The findings are self-reported, not verified incident…

Employee AI misuse rivals phishing in security incident survey
Summary
  • In IO’s survey, approximately 25% of respondents who experienced an incident or near-miss cited one of two types of employee AI misuse as the leading cause of their most serious case.
  • The comparison figure for phishing was 24%, while 35% of all surveyed managers identified unapproved generative AI use as a common staff mistake.
  • Although 51% said their organisations had updated AI policies, respondents also reported unapproved AI workflows and disclosure of sensitive information.

Employee misuse of artificial intelligence featured as prominently as phishing in a new survey of UK and US cybersecurity managers, with respondents who had experienced an incident or near-miss frequently citing unauthorised AI activity as the leading cause of their most serious case. The research, commissioned by business resilience provider IO, describes managers’ reported experience of employees connecting AI tools to business systems or sharing information without authorisation.

IO reports that approximately 25% of respondents who had experienced a cybersecurity or information security incident or near-miss in the previous year attributed their most significant case to one of two AI-related staff behaviours. Around 14.45% cited giving an AI tool access to a business system without approval, and 10.33% cited leaking confidential information through an AI tool without approval. The comparable figure attributed to phishing was 24%. These percentages describe answers from an affected subset of surveyed managers, not measured shares of all cyber incidents or of all businesses.

The survey was carried out by Censuswide among 3,000 cybersecurity managers and more senior respondents, comprising 2,000 in the UK and 1,000 in the US. Fieldwork took place between 14 September and 1 October 2026 across several sectors, including finance, healthcare, manufacturing, utilities and professional services. Its scale allows comparison of managers’ reported experience, but the results do not establish incident causation through independent forensic examination or prove that the same pattern applies to every organisation.

In a separate question about employee behaviour, 35% of respondents cited use of generative AI without permission or guidance as a common security mistake, ahead of 32% identifying clicks on suspicious links. That comparison concerns the frequency with which managers recognise particular behaviours, rather than a count of successful intrusions caused by each one. It nevertheless indicates that governance of routine workplace AI has become an operational concern alongside established credential theft and social engineering problems.

Unapproved AI use can create several different exposure paths. An employee may paste confidential customer information into a public service, connect an AI agent to a repository with broad permissions or automate an internal process without documenting the data it accesses. These activities do not all constitute confirmed breaches, and some may violate policy without creating a reportable event. Their common feature is the difficulty of identifying where data travels and who has authorised an automated action.

IO says 22% of respondents reported staff incorporating AI into important workflows without the security team knowing, while a similar share reported confidential or sensitive information being shared with public generative AI tools. The release describes these behaviours as evidence of limited visibility, but the underlying survey cannot demonstrate how much data was disclosed in each case or the downstream consequences. The operational risk depends on the nature of the information, the service’s handling rules and the level of integration granted.

Formal policy adoption has not eliminated the reported visibility problem. Just over half of respondents, 51%, said their organisation had updated its AI usage policies or guidance. Among organisations with at least 250 employees, 46% reported an enforced formal AI policy, compared with 36% among those with fewer than 100 staff. Yet reports of important AI workflows operating without security-team knowledge were 24% and 21% respectively. Those comparisons are descriptive: they do not prove that policies are ineffective or identify the impact of a particular control.

Organisations can struggle with accountability when AI is introduced through everyday business applications rather than a centrally procured platform. Security teams may oversee access and information handling while operational managers choose tools and employees decide how to use them. A policy written by one function does not automatically give that function visibility into every data connection. Governance therefore involves roles, approvals and technical controls as well as employee training, although the study does not test which combinations reduce incidents.

The supplier also reports disagreement over responsibility for resilience, with 34% of respondents identifying the chief executive as ultimately accountable and 27% saying their organisation is shifting AI security accountability from the security function to individuals. Those findings speak to governance arrangements and reported perceptions, not to a statutory allocation of liability. Regulatory obligations, employment responsibilities and contractual duties can differ by jurisdiction, processing activity and the role of the organisation concerned.

IO argues that standards addressing information security, data privacy and AI management offer complementary governance frameworks. As a supplier of compliance and resilience software, IO has a commercial interest in stronger AI governance practices; the survey does not independently test the effectiveness of its preferred controls. In the affected subgroup, the combined share attributing the most serious incident or near-miss to employee AI misuse was approximately 25%, compared with 24% for phishing. Further incident-level evidence would be required to establish whether this perceived pattern reflects a sustained change in actual attack and data-loss causes.

×