AI at work
Brazil ANPD tests employee AI tools in third sandbox report
The Brazilian data protection authority published its third AI sandbox report, highlighting the gap between declared and verified AI security controls.
On October 16, 2026, Brazil's National Data Protection Authority (opens in a new tab) (ANPD) published the third partial monitoring report of its regulatory sandbox for artificial intelligence and data protection. The document outlines the results of security and privacy tests conducted on three technological solutions. The sandbox operates as an experimental, controlled, and supervised environment, run in partnership with the Center for Artificial Intelligence and Machine Learning at the University of São Paulo.
While the initiative is experimental and does not attest to the legal compliance of the tested solutions, the report provides insight into how the regulator evaluates artificial intelligence systems. A central theme of the third cycle was the transition from reviewing documentation to verifying actual security controls. For organizations that deploy artificial intelligence tools for their employees, the findings emphasize the importance of validating vendor claims, setting clear usage boundaries for professionals, and understanding exactly how these systems process personal data.
Who it applies to
The sandbox directly involves three companies selected through a public notice issued by the technological innovation superintendence in 2025. The participants are Metatext, operating in the financial sector; Synapse AI, in mobility; and Prevvine Tecnologia, in healthcare.
The authority explicitly states that the sandbox is aimed at learning and producing evidence. It does not limit the actions of the participating companies, nor does it produce enforcement or fiscalization effects. However, the authority noted that the knowledge generated will subsidize future orientations and norms.
For the broader market, the report serves as a practical indicator of the areas the regulator prioritizes. Any company that acts as a controller and provides third-party artificial intelligence tools to its workforce can observe the technical and organizational controls evaluated in this cycle. The tests focused heavily on the treatment of personal data related to financial operations, location, and health information, which are common data types handled by employee-facing applications.
What it requires
Because the sandbox is a testing environment, the third report does not impose new legal obligations on the general public. Instead, it details the specific conditions the participating companies must meet to proceed to the next testing cycle.
According to the authority, the third cycle marked a maturation in the testing process. The evaluations considered the technological maturity, technical particularities, and application sectors of each solution. A major development was the shift from methodological structuring and documentary analysis to the active verification of security mechanisms. The authority highlighted the necessity of distinguishing the information declared by the companies from the information that is documented and effectively verified. For companies buying artificial intelligence tools for their employees, this means vendor marketing materials and initial privacy declarations are insufficient. The actual technical safeguards must be validated.
In the healthcare sector, the sandbox evaluated a solution that offers clinical decision support to medical professionals. The tests utilized 30 simulated clinical cases to analyze the system responses. To continue in the sandbox, the authority set conditions requiring the tool to present clear warnings about its usage limits, particularly in emergency situations. The system must also clearly indicate the sources consulted to elaborate its answers. This ensures that professionals relying on artificial intelligence for their daily tasks can verify the accuracy of the outputs before making decisions.
In the financial sector, the work concentrated on analyzing the documentation and data flows of a system that uses artificial intelligence to detect fraud and block suspicious operations. For the next cycle, the tests will focus on personal data protection mechanisms and the emission of alerts. The authority will also test ways to explain automated decisions to the individuals who have their operations blocked. Organizations deploying similar tools for their internal finance or compliance teams must ensure the systems can provide clear reasoning for their outputs.
In the mobility sector, the tests evaluated an application that recommends routes based on user location. During simulated use, the authority monitored data traffic and found no personal information circulating without protection. The tests also verified that the algorithm used to estimate the concentration of people in images did not identify individuals. The next stages will evaluate accessibility aspects and how the application shares data with third-party services, such as map platforms. This highlights the need for companies to maintain strict inventories of where their employee tools send data.
The sandbox operates as a collaborative environment. During the activities, the organizations share information about their technical characteristics and governance structures. The authority and the university work together to identify alternatives for improvement when facing potential regulatory nonconformities.
The dates
The National Data Protection Authority published the third partial monitoring report on October 16, 2026.
The participating companies are now entering the fourth monitoring cycle. In this upcoming phase, the organizations must execute the planned tests and prove they have fulfilled the conditions identified in the previous stages.
The authority has not announced a specific date for the publication of final binding norms derived from this project. The results from the fourth cycle will continue to contribute to the production of evidence for future regulations.
What to check now
- Inventory the artificial intelligence tools that employees use for clinical, financial, or operational decisions.
- Compare the security and data protection controls declared by the artificial intelligence vendors against the verified technical documentation.
- Check if the artificial intelligence tools used by the workforce provide clear and visible warnings about their operational limits.
- Ensure that systems generating professional recommendations or answers display the exact sources consulted.
- Establish a documented process to explain automated decisions to individuals affected by fraud detection or blocking mechanisms.
- Map whether internal artificial intelligence applications share data with third-party services and verify the legal basis for that transfer.
- Assess whether algorithms used for general estimations inadvertently identify specific individuals during their data processing routines.
Sources
This article is general information, not legal advice. For your company's situation, talk to a qualified lawyer.
This article was written with AI and checked, sentence by sentence, against the sources linked above.