OWASP Evolution
Theme

Data and Model PoisoningLLM05: Data and Model Poisoning · 2026

Tampered data or artifacts persistently corrupt learning, embeddings, RAG or continuous feedback and can conceal backdoors.

Summary

Tampered data or artifacts persistently corrupt learning, embeddings, RAG or continuous feedback and can conceal backdoors.

How the attack works

LLM05 · How the attack worksTampered training data teaches a false association; the model repeats that answer after fine-tuning.Clean dataTampered dataFine-tuningModel: false association

Tampered training data teaches a false association; the model repeats that answer after fine-tuning.

Example

Vulnerable · Pseudocode
dataset += unreviewed_feedback
model = fine_tune(base, dataset)
# Poisoned labels affect future answers

Fix: Version and verify data provenance; isolate unreviewed contributions and evaluate anomalies before and after fine-tuning.

Prevention

  • Maintain lineage, signatures, hashes and recoverable versions.
  • Restrict data injection and supervise retraining.
  • Monitor drift, anomalies and triggers.

Lineage

  1. 2025 · LLM04: Data and Model Poisoning Moves down to rank 5 in 2026
  2. 2026 · LLM05: Data and Model Poisoning Moves down from rank 4
  • 2025:LLM04 → 2026:LLM05: Data and Model Poisoning moves down from rank 4 to 5.

Sources

Buy me a coffee

OWASP Evolution

Editions
Continues / movesContinues RenamedRenamed Merged / expandedMerged