One out of every 28 queries sent to generative AI tools from corporate environments presents a high risk of sensitive information leakage. This data appears in a report by Check Point Research and directly affects companies that have already incorporated these systems into their daily work.
The main paradox is that the implementation is already massive, but the risk is also massive. The study indicates that 91% of organizations using generative AI in their daily operations have been affected by this danger of exposure, both on a general scale and in the specific case of Spain.
Spain registered 1,883 weekly cyberattacks per company in December 2025
The context in which corporate AI use is growing was already one of sustained pressure on cybersecurity. In Spain, companies suffered an average of 1,883 weekly cyberattacks in December 2025, a 5% increase from the previous year.
The daily use of generative assistants and platforms is now superimposed on this scenario. The report indicates that 91% of Spanish organizations using generative AI faced high-risk queries, a category that includes interactions with the potential to expose information that should not leave the internal perimeter.
The most compromised data in these types of queries include credentials and internal secrets, personal data, financial information, intellectual property, and commercial information. It's not just about complete documents, but also about code snippets, access keys, internal figures, or operational details entered by employees in prompts.
Samsung, Apple, and Amazon banned ChatGPT after internal leaks
Some major tech companies had already adopted restrictions after specific incidents. Samsung, Apple, and Amazon banned the internal use of ChatGPT after detecting leaks of source code and other confidential data by employees.
This precedent illustrates a problem that is no longer limited to the visible use of a specific application. Check Point Research adds that nearly 40% of MCP servers, an acronym for Model Context Protocol, presented security vulnerabilities in the sample examined.
MCP servers act as a connection point between models and external sources of information or services. If this layer fails, the exposure depends not only on what an employee writes in a prompt but also on the permissions, accesses, and integrations surrounding the tool.
Companies concentrate risk in permits, access, and data classification
The report proposes several measures to reduce exposure in corporate environments. The first is to define an internal policy for AI use that specifies what data can be entered, into which tools, and for what tasks.
Along with that rule, it recommends organizing information into four classification levels. It also proposes activating SSO and MFA systems on the platforms used, reviewing integration permissions, and establishing an incident response procedure.
Outside of Spain, the same regional analysis records a 15% increase in cyberattacks in Argentina. In parallel, the technical review by Check Point Research detected security vulnerabilities in approximately 40% of the MCP servers examined.