A recently off‑boarded contractor leaves an OpenAI‑powered automation script running in a CI pipeline. The script continues to call the OpenAI Agents SDK with a hard‑coded API key, pulls customer data from internal services, and writes results to a shared bucket. No one sees the request, no log captures the exact query, and the data that flows through the model is never inspected or masked. In a jurisdiction that enforces the Lei Geral de Proteção de Dados (LGPD), that silence becomes a compliance nightmare.
What LGPD demands from AI‑driven workloads
LGPD focuses on three pillars that directly affect any system that processes personal data: accountability, data minimisation, and the right to audit. Accountability requires organisations to demonstrate who accessed what data, when, and under which authority. Data minimisation obliges you to limit the amount of personal information exposed to downstream services, including large language models. Finally, the law requires auditors to be able to retrieve a reliable record of every processing event, and that the data subject can request a copy of the processing log.
Why the OpenAI Agents SDK alone cannot satisfy those requirements
In practice, teams often embed the SDK in scripts, bots, or serverless functions that authenticate with a static secret. The secret is stored in code repositories or environment variables, and the SDK forwards requests straight to the OpenAI endpoint. This architecture provides no point where a policy engine can inspect the payload, enforce masking, or require a human approval step. The request travels directly from the runtime to the model, leaving a gap where LGPD‑required evidence cannot be captured. Even if you add an OIDC identity layer, the token only proves who started the job; it does not record the exact query or the response that may contain personal data.
hoop.dev as the data‑path enforcement layer
hoop.dev inserts a Layer 7 gateway between the OpenAI Agents SDK and the model endpoint. By placing the gateway in the data path, hoop.dev becomes the only place where enforcement can happen. It records every session, retains a complete audit trail, and can apply inline masking to any fields that match personal‑data patterns before the response reaches the downstream system. Because hoop.dev mediates the connection, it can also require just‑in‑time approval for high‑risk queries, ensuring that a privileged user signs off before the model processes sensitive identifiers.
Session recording for accountability
hoop.dev captures the full request and response pair for each SDK call. Those logs are stored outside the runtime that generated the request, providing auditors with reliable evidence that satisfies LGPD’s accountability clause. The logs include the identity of the caller, the timestamp, and the policy decision that was applied.