Best practiceGovernance and people
Secure the data your AI is trained on and reads
OWASP LLM04OWASP LLM02OWASP LLM08
What a model learns and what an assistant retrieves can only be as trustworthy as the data behind it, and a leak of that data is a leak of everything in it: DeepSeek left a database of chat history open to anyone. Joint guidance from CISA, the NSA, the FBI and partner agencies names three risks across the AI lifecycle: the data supply chain, maliciously modified data, and data drift.
What to do
- Track where every dataset and document comes from and log the path it takes through the system (its lineage), so a poisoned or leaked source can be traced.
- Check integrity with checksums or cryptographic hashes in storage and in transit, and sign the revisions you trust.
- Classify data by sensitivity, and limit which people and agents can read each class.
- Encrypt it at rest and in transit, and keep it on infrastructure you trust.
- Mask personal data before training or indexing, or use privacy-preserving methods such as differential privacy or federated learning.
- Delete data securely when it is no longer needed, and reassess its risks on a schedule.