Not test data. Real data. The kind that’s supposed to stay locked behind your walls forever. It happened fast—so fast you almost missed it—but it was enough to make you rethink every assumption about how you govern generative AI.
Generative AI can draft code, answer questions, and build artifacts in seconds. That same speed can leak sensitive information just as quickly if you don’t have tight data controls in place. The NIST Cybersecurity Framework offers a proven structure for managing risk. The challenge is adapting it to handle the unique risks of large language models and other generative systems.
The core of the NIST Cybersecurity Framework—Identify, Protect, Detect, Respond, Recover—maps cleanly to AI safety, but requires a shift in focus: protecting prompts, training sets, embeddings, and generated outputs.
Identify every data source your AI can access. That means mapping model inputs, outputs, and hidden connections to services where data may be stored or cached. Threat modeling is no longer optional; it must include model behavior under attack, data poisoning, and prompt injection.
Protect by implementing strict access controls for model interactions. Sanitize prompts and strip sensitive fields before they reach the model. Encrypt stored embeddings and limit retention to only what is necessary. Apply differential privacy or redaction where possible to prevent confirmations of sensitive facts.