On October 16, 2026, Brazil’s National Data Protection Authority (ANPD) reported different findings for three artificial intelligence solutions in its regulatory sandbox: in mobility, it checked data traffic safeguards and image-processing tests; in healthcare, it documented 30 simulated clinical cases and problems with references; and in the project linked to finance, it reviewed documentation but deferred technical testing to the following cycle. Participation does not amount to certification, clinical approval or proof of full compliance with Brazil’s General Data Protection Law (LGPD).
The Third Partial Monitoring Report covers Cycle 3 of the pilot run by ANPD with the University of São Paulo’s Center for Artificial Intelligence and Machine Learning (CIAAM/USP). The participants are Synapse AI, Prevvine Tecnologia and Metatext. The report itself distinguishes what was tested, what was submitted in documents and what still lacks practical verification.
What the regulatory sandbox actually assesses
A regulatory sandbox is a supervised experimental setting for monitoring risks and testing safeguards under defined conditions. The methodology specifies six cycles; from the second to the sixth, each includes preparation, experimentation and risk management, followed by evaluation and adjustments. The purpose is to build evidence for future regulatory decisions, not to issue an operating licence or declare every product safe in all circumstances.
The assessment combines documents, records and technical tests with an ethics framework worth up to 100 points, based on UNESCO recommendations. The score describes the criteria assessed in that cycle, including governance, transparency and security. It is not an algorithmic accuracy rate or a clinical effectiveness score. Nor does it establish compliance with every legal duty.
Three projects, three different levels of evidence
| Sector and solution | Documented or verified in Cycle 3 | Still to be demonstrated |
|---|---|---|
| Mobility: Synapse AI, Trajetto | Audited traffic showed encryption in transit; image processing did not identify individuals under the test conditions. | Third-party service flows, indirect identification risk, accessibility, user explanations and application security. |
| Healthcare: Prevvine, STAiDOC | HarpIA integration restored, 30 simulated clinical vignettes applied and failures identified in bibliographic references. | Consolidated clinical results, transparency, personal data protection, security and warnings about use limitations. |
| Finance and AI security: Metatext, GuardionAI | Architecture, data flows and regulatory documents reviewed; two tests planned for the next cycle. | Practical data protection and alert tests, human oversight and clarification of international data transfers. |
Mobility: what Trajetto testing actually verified
Trajetto uses journey information and estimates of pedestrian flows. Researchers audited communications between a modified app and a database in an environment that captured traffic and used synthetic data. In the observed flow, personal data was encrypted in transit and user identifiers were not exposed in clear text. In a separate test, the crowd-concentration algorithm did not identify individuals in the synthetic images examined.
These findings have concrete limits. The capture did not necessarily cover all traffic or calls to third-party services, such as mapping platforms. Even without receiving someone’s name, an outside provider may, under certain conditions, link a request to a device or a user’s location. Accessibility, location permission and transparency issues also remained. Synapse’s ethics score rose from 45 to 68 points, with high inherent risk, medium residual risk and an assessment that allowed continuation subject to conditions.
Healthcare: 30 simulated cases do not validate clinical accuracy
STAiDOC supports healthcare professionals’ decisions. Cycle 3 restored its connection to HarpIA, the environment used for evaluations, enabling the application of 30 clinical vignettes, meaning simulated case descriptions with variations in patient profiles and context. However, the detailed report explains that the case records contain no consolidated pass or fail results for the responses. There were also temporary integration problems. The 30 cases therefore do not prove diagnostic accuracy, clinical safety or the absence of bias.
Reviewers found hallucinated answers, nonexistent references and broken links in the NTAV bibliographic verification component. Its technology readiness level was revised from TRL 7 to TRL 6; TRL measures technological maturity, not medical effectiveness. Prevvine’s overall ethics score rose from 60 to 63 points, with critical inherent risk, medium residual risk and a favourable assessment subject to conditions. Outstanding requirements include displaying sources, warning against use in emergencies, providing a visible channel for human review, protecting against re-identification, and demonstrating application security and service continuity.
The report also recommends checking whether the solution’s functions require classification as software as a medical device. That is a pending sector-specific regulatory assessment, not a completed classification or approval by Brazil’s health regulator Anvisa.
Finance: why GuardionAI has no completed technical test result
ANPD’s public announcement places Metatext’s project in the financial context and mentions fraud detection and blocking suspicious transactions. The more specific technical report identifies GuardionAI as a platform to secure and control AI-agent and generative AI systems. In Cycle 3, the work was primarily documentary, examining architecture, data flows and potentially applicable obligations. There is no demonstration that ANPD tested and validated real-world financial transaction blocking.
The regulatory mapping expanded from 5 to 58 legal instruments, a broader review rather than 58 confirmed instances of compliance. Two independent tests were designed for Cycle 4: one on personal data protection or pseudonymisation, the other on alerts sent to the security team. Open questions concern international data transfers, traceability, responsibilities, human oversight and explainability of system decisions.
What an assessment subject to conditions means
The report separates safeguards claimed by a company, those described in documents, implemented mechanisms and mechanisms actually observed in tests. Trajetto and STAiDOC were allowed to continue experimental work subject to conditions. This is not definitive compliance, a fair ranking of systems serving different purposes or unrestricted permission to process personal data. GuardionAI’s central technical checks remain scheduled for the next cycle.
For Cycle 4, ANPD calls for evidence of protection in external integrations and accessibility in mobility; analysis of vignette results, references and warnings in healthcare; and execution of GuardionAI’s outstanding tests. These lessons may inform future guidance, but the report did not introduce a new general AI rule and does not replace existing LGPD obligations or sector regulators’ responsibilities.
