Data and model poisoning introduces malicious influence into training, fine-tuning, or retrieval corpora to bias outputs or implant hidden triggers. Attackers can shape responses without direct system compromise by contaminating trusted data paths. Detection is difficult because effects may appear gradual or context-specific.
Implement provenance controls, integrity validation, and anomaly screening across all data ingestion and training workflows. Segment trusted and untrusted corpora and apply quarantine processes for suspicious samples. Continuously evaluate model behavior for drift and trigger patterns using adversarial testing.
The same feature written the risky way, then with the fix applied.
How this risk has ranked in each OWASP release.
2025Latest
#04llm04-2025 Data and Model Poisoning
