A newly offboarded contractor still has a service account that CrewAI uses to pull customer records for a nightly report. The account can query any table, export raw rows, and write results to a shared bucket. When the contractor leaves, the token remains active for weeks, and no one knows which queries were run or what personal data left the environment.
Under Brazil’s General Data Protection Law (lgpd), organizations must demonstrate accountability for every access to personal data, keep tamper‑evident logs, and ensure that only the minimum necessary information is exposed. Auditors look for concrete evidence: who accessed what, when, and whether the data was masked or approved before leaving the system.
For AI‑driven platforms like CrewAI, the challenge is twofold. First, the model often runs as an automated job that authenticates with a static credential. Second, the job can issue ad‑hoc queries that bypass human review, making it hard to prove compliance after the fact.
What lgpd expects from AI‑driven data pipelines
lgpd requires that personal data processing be transparent, auditable, and limited to the purpose originally disclosed. Key evidence points include:
- Identity of the entity (human or service account) that initiated a request.
- Exact query or command executed against the data store.
- Timestamp and duration of the session.
- Any transformation or masking applied to protected fields.
- Approval records for operations that exceed a predefined risk threshold.
Without these artifacts, regulators consider the organization non‑compliant, and penalties can be severe.
Where the gap appears in typical CrewAI deployments
Most teams give CrewAI a long‑lived database credential and let the model connect directly to PostgreSQL, MySQL, or another datastore. The credential is stored in a CI secret manager, but the connection bypasses any gateway that could inspect the traffic. As a result:
- There is no per‑query audit log; only the database’s generic connection logs exist.
- Sensitive columns such as CPF, email, or address are returned in clear text.
- High‑risk queries (e.g., full table scans or export commands) execute without any human sign‑off.
- Session replay is impossible because the database does not retain the command stream.
These gaps mean that even if the organization implements strong identity‑provider policies, the enforcement outcomes required by lgpd never materialize.
How hoop.dev provides the missing controls
hoop.dev is a layer‑7 gateway that sits between CrewAI and the target infrastructure. By proxying every connection, hoop.dev becomes the only place where enforcement can happen.