All threats

ai-training-data-poisoning

Attacker manipulates training data or fine-tuning datasets to introduce backdoors or biases into AI models

MITRE ATT&CK techniques

Common Weakness Enumeration

• CWE-20

Mitigating controls

ctrl-poison-1

Validate and sanitize all training data sources

ctrl-poison-2

Implement data provenance tracking

ctrl-poison-3

Use anomaly detection on training datasets

ctrl-poison-4

Conduct regular model behavior audits

ctrl-poison-5

Implement access controls and audit logging on training data storage

ctrl-poison-6

Verify training data integrity using checksums or cryptographic hashes before training runs

References

• https://genai.owasp.org/llm-top-10/
• https://atlas.mitre.org/