GuardDuty RDS: detect exfiltration without mistaking detection for governance
AWS has expanded GuardDuty RDS Protection to detect behaviours associated with data exfiltration and destruction on Amazon Aurora PostgreSQL and Amazon RDS for PostgreSQL. This layer is useful when valid credentials have already been compromised; it does not make a business database governed by default.
1. What the announcement confirms
AWS says “RDS Protection for data activity” establishes a baseline of usual queries and combines anomaly detection with threat intelligence. It supplements login protection and can be enabled organisation-wide or through APIs, SDKs, the CLI, or CloudFormation for supported PostgreSQL deployments.
2. Why a legitimate identity remains a risk
A SQL query run by an authorised account can look normal until its volume, origin, or pace diverges from the baseline. Phishing, social engineering, or misconfiguration can therefore provide access that login controls alone cannot fully qualify. An alert improves visibility; it does not itself prove exfiltration or cover every data path.
3. What changes for a Belgian or French company
ERP, CRM, billing, and support data may feed RAG, an agent, or Odoo Enterprise automation. Before granting access, separate human, application, and agent identities; limit tables and operations; and log reads as well as writes. SMEs should start with a non-critical database and read-only roles. Mid-market firms, large enterprises, and public administrations must connect alerts to on-call ownership, a data owner, and a triage procedure.
4. Underside analysis: an alert must lead to a response
Teams need to understand the query, isolate the identity, suspend integrations, preserve evidence, and restore from a tested backup. Rights management, secret rotation, encryption, rate limits, and network segmentation remain necessary. A European architecture does not remove the need to verify residency, support access, retention, and log export.
Operational recommendation: for every AI-connected database, map identities, tables, tools, expected volumes, and business owners. Simulate unusual extraction using test data, measure alerting, revocation, and recovery, then—and only then—enable Odoo or ERP writes.
Secure data in an AI architecture