NSA, CISA Urge Organizations to Secure Data Used in AI Models

The NSA, CISA, and FBI jointly issued guidance urging organizations to secure data used to train and operate AI systems by addressing supply-chain risks, maliciously modified or "positioned" data, and data drift. The guidance lists ten best practices—including provenance tracking, data integrity verification, cryptographic signatures (quantum-resistant where possible), zero-trust/trusted infrastructure, classification and access controls, AES-256 encryption, certified secure storage, privacy-preserving techniques (differential privacy, federated learning), secure deletion, and continuous data security risk assessments—and emphasizes that implementation may require investment and ongoing monitoring to mitigate LLM-specific risks like data poisoning.