In May 2026, two lawyers were fined R$84,000 by the Regional Labor Court of the 8th Region (TRT-8) for manipulating the court’s AI system, “Galileu”. Using the prompt injection technique, they hid commands in white text in a petition so that the system would carry out only a superficial analysis. This episode exposes the fragility caused by the absence of policies on the use of these technologies and of adequate controls over how they operate. Shall we take a closer look at this controversy?
1. BEYOND DOCUMENTS: WHY DON’T AI POLICIES GUARANTEE SECURITY ON THEIR OWN?
The corporate race to adopt generative AI has moved much faster than organizations’ capacity to oversee it. Under pressure to take an institutional position, many companies chose an exclusively documentary path. This strategic layer sets the “rules of the game”, but it may operate only at the level of intention. In this context, governance is treated as an end in itself rather than as a dynamic process.
The problem is that these documents regulate the intended use, not the model’s actual behavior. For example, a policy may prohibit employees from entering confidential data into generative AI tools, but if access to the tool has no technical lock (such as blocking uploads of certain file types), the prohibition may not be effective. In other words, the model continues to accept any input, since it is not guided by the company’s code of conduct. It only reads configuration parameters, allowlists, system rules and moderation layers.
This disconnect between the written rule and technical control is what is known as the implementation gap, in which the mismatch between technological progress and risk management methods can be catastrophic. Operating AI tools without technical control mechanisms generates costs that go beyond the financial, reaching the reputational and structural spheres of organizations. Because the behavior of generative models is not entirely predictable, governance must evolve from static documents into a continuous ecosystem, in which technology works in accordance with the chosen rule. Security only emerges from the convergence between guidelines and policies and the proper configuration of tools.
2. AI GOVERNANCE IN PRACTICE: FROM THE STRATEGIC PLAN TO TECHNICAL CONFIGURATION
Proper governance requires that the “rules of the game” defined at the strategic level (policies and guidelines) be translated into operational controls at the technological level (tool configurations). In other words, a code of conduct is not enough; it takes a multidisciplinary ecosystem in which legal, information security and data science teams collaborate to monitor the full life cycle of AI technologies. This convergence ensures that technology operates within ethical and legal limits, turning documents into auditable, traceable and continuous processes.
Part of this practice lies in the technical management of the information assets that feed artificial intelligence. Some essential examples for structuring this architecture include:
- Data Classification: labeling information according to its sensitivity and value, enabling differentiated protection policies and access controls.
- Mapping and Lineage (Data Lineage): tracking the flow, origin and transformations of data to identify vulnerabilities and ensure that the “raw material” is sound and lawful.
- Role-Based Access Controls (RBAC): ensuring that only authorized profiles (such as specific developers or risk analysts) access subsets of sensitive data, mitigating the risk of leaks.
- Use of Privacy Enhancing Technologies (PETs): applying Privacy by Design through techniques such as homomorphic encryption or synthetic data, enabling information to be processed without exposing the underlying data.
- Continuous Real-Time Validation: establishing constant monitoring mechanisms to track unforeseen behavior, hallucinations or value drift during dynamic GenAI interactions.
At the operations and security layer, governance materializes as systemic locks that mitigate risks. This involves adopting Model Cards, which document in a standardized way the capabilities, limitations and biases of models for users and auditors. It is also essential to carry out prompt injection tests and implement content filters to block malicious manipulation and inappropriate outputs. Finally, human oversight (human-in-the-loop) in high-impact decisions ensures that the system’s autonomy does not remove the organization’s ultimate responsibility for the results it generates.
3. THE COST OF NEGLIGENCE: FINANCIAL, REPUTATIONAL AND STRUCTURAL IMPACTS
Operating AI systems without adequate technical governance generates costs that go beyond immediate losses, affecting the organization’s long-term viability. There are three main categories of costs:
- Financial Costs: these are direct monetary losses, including heavy fines imposed by regulators or courts, high litigation costs and the loss of valuable assets due to fraud or systemic failures. Mitigating them requires a solid compliance and regulatory strategy, with continuous monitoring of data quality and technical validation of models before and during operation. One example is the R$84,000 fine imposed by TRT-8 following a prompt injection attack, which the court’s purely documentary governance was unable to prevent technically and proactively.
- Reputational Costs: this is damage to the institution’s image and loss of trust among users and the market. Discriminatory or opaque automated decisions destroy brand value and can lead to humanitarian and political crises. Avoiding them requires transparency practices, such as using Model Cards to document limitations and data origins. One example is the childcare benefits case in the Netherlands, where the use of algorithms to detect fraud led to thousands of families being falsely accused on the basis of their racial profile, triggering a crisis in the Dutch government.
- Structural Costs: these involve the creation of critical systemic vulnerabilities, high-maintenance technical debt and inefficient models operating outside the organization’s control. They also manifest as a loss of control over information flows and exposure to sophisticated cyberattacks. Implementing role-based access controls (RBAC), data mapping and lineage, and the use of PETs are essential to mitigate them. An example can be seen in the case of the leak of 150 GB of data from the Mexican government. On that occasion, a criminal exploited the Claude chatbot to act as an insider in government networks and steal records of 195 million taxpayers, as well as employee credentials.
4. A PARADIGM SHIFT: AI GOVERNANCE AS A CONTINUOUS ECOSYSTEM
The incidents discussed reinforce that AI governance cannot be treated as a static project, but rather as a continuous ecosystem in which monitoring and technical assessment evolve alongside the technology. This paradigm shift requires that the documentary layer, where the “rules of the game” are set, and the actual configuration of tools operate in full convergence to avoid security gaps.
This is precisely the debate that Data Privacy Brasil brings to the Data Privacy Global Conference (DPGC 2026): how can those who write the policy and those who configure the system be brought closer together in practice? And why has this connection become so relevant? Anyone who wants to explore this discussion further is invited. See you at DPGC 2026!




