LLM04: Data and Model Poisoning · 2025

Tampering with data, embeddings or model artifacts introduces bias, degradation or backdoors triggered by specific conditions.

Summary

Tampering with data, embeddings or model artifacts introduces bias, degradation or backdoors triggered by specific conditions.

How the attack works

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

Record provenance and versions of data and models.
Isolate unverified sources and detect anomalies.
Monitor behavior and run robustness tests.

Lineage

2025 · LLM04: Data and Model Poisoning Moves down to rank 5 in 2026
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

Open official source ↗

Buy me a coffee