Training data poisoning manipulates datasets to influence model behavior with targeted bias, hidden triggers, or malicious latent behaviors. In retrieval-augmented and fine-tuned systems, poisoned corpora can silently shape outputs and degrade trust. Because poisoning may remain dormant, detection requires lifecycle monitoring and provenance controls.
Harden data ingestion pipelines with source verification, integrity checks, and quarantine workflows for suspicious inputs. Track provenance and change history for training and retrieval corpora. Monitor behavior drift and run red-team evaluations to detect triggerable or anomalous model responses over time.
The same feature written the risky way, then with the fix applied.
How this risk has ranked in each OWASP release.
2024Latest
#03llm03-2024 Training Data Poisoning
