Healthcare Employees Sharing Patient Data with AI Tools Viewed as Major Emerging Risk
Healthcare security leaders are increasingly concerned about patient privacy incidents, HIPAA violations, and intellectual property loss due to employees sharing sensitive data with public AI models. While there is a risk of employees uploading patient data to unapproved tools such as ChatGPT, Gemini, or Claude, sensitive data may also be disclosed by unauthorized browser extensions, plug-ins, or an OAuth grant, which can be approved in a few clicks. According to a recent survey by the cybersecurity company DNS Filter, 43% of healthcare security leaders said they believe employee sharing of sensitive data with AI models is the biggest emerging AI risk in the next 12 months. The biggest perceived risk is loss of intellectual property through AI training data, rated as the biggest risk by 53% of respondents, compared to 31% in other sectors.
Multiple privacy incidents have already been reported by healthcare providers involving employees uploading sensitive patient data to public AI models, in violation of their organization’s data privacy and security policies and the HIPAA Rules. Recent examples include Johns Hopkins Bayview Medical Center, Weill Cornell Medicine, and the Washington Department of Social and Health Services. Since breaches of protected health information are only made public via the HHS’ Office for Civil Rights breach portal if they affect 500 or more individuals, and media notices are only required for breaches of that magnitude, the problem could be far more extensive. The DNSFilter survey findings suggest that may well be the case.
Around half of healthcare respondents have either confirmed (27%) or suspect (23%) that employees have connected an unapproved AI tool to Google Workspace, Microsoft 365, or their SSO environment. While patient data may not have been involved in those actual or suspected incidents, it is a significant risk, and a lack of AI governance means privacy violations may go unnoticed.
According to the DNSFilter report – Seeing the Risk, Missing the Signal: AI Security in Healthcare – only 27% of respondents said every security tool they own is actively monitored, and 40% of respondents said they actively monitor 50% or fewer security tools. While 70% of organizations confirmed they have implemented formal policies prohibiting the use of unapproved AI tools, only 40% said their policy is consistently enforced. Half of respondents said AI security and governance is one of their biggest security gaps; the highest share of any sector.
Six out of 10 respondents said they are reliant on their AI and SaaS vendors to alert them to security incidents, which is more than the 43% of respondents from other sectors. Only 23% of healthcare respondents said they catch SaaS and AI incidents through employee reporting, compared to 44% in other sectors. An additional concern is that 47% of healthcare organizations said their employer can authorize a new AI tool to access company systems without higher-level approval.
Despite the security gaps, confidence is high in the ability to detect breaches. In healthcare, 87% of respondents said they are confident that they could detect a breach within 6 hours, and 40% said they are very confident. One-third said they had experienced an incident that originated from a SaaS or AI tool in the past 12 months, and the same percentage said they had experienced such an incident that was never formally reported or disclosed.
Despite the risks, only a small percentage of healthcare organizations rate AI security and governance as one of the most important areas for IT security spending. Only 33% of respondents said they are sending their next dollar to AI security and governance, compared to 41% in other industries. This shows that while concern about risk is growing, money to address the problem is not following.
“Healthcare security leaders know exactly where AI puts their data at risk. What they’re missing is the moment that risk becomes an incident, and too many are hearing about it from a vendor instead of their own team. When the data involved is protected health information, the risk only becomes more significant. Before the next dollar goes to AI governance, healthcare organizations need to see which AI services their people are already using without clearance,” said Ken Carnesi, CEO, DNSFilter.
