Security protects sensitive data, models, infrastructure, communications, storage, logs, and analytical operations from unauthorized exposure or unsafe change. Generative-AI workflows also require clear responsibility for model modifications, external providers, prompt histories, and regulatory compliance. section 8.6
Use these checks when evaluating an AI model or analytics workflow:
- Document retraining, fine-tuning, and other model changes; unauthorized changes can invalidate results or introduce bias. section 8.6
- Encrypt communication with externally hosted models and agree explicitly on how prompts, responses, and logs are stored or reused. section 8.6
- Separate contexts between users and sessions where possible, especially when a provider retains prompt history or model memory. section 8.6
- Treat sensitive data shared with third-party models as a potential exposure, and assess the provider’s security responsibilities. section 8.6
- Prefer verified models, inspect documentation and contributors, scan for malicious code, isolate testing, limit permissions, update dependencies, and encrypt sensitive data. section 6.2.4
Evaluating a downloaded model safely
Before connecting a model to sensitive records, verify its source, review its documentation and contributor history, scan it for vulnerabilities, and test it in an isolated environment. After approval, grant only the permissions it needs, keep dependencies updated, and encrypt sensitive data. Recheck community reports and security guidance as the model and its environment change. section 6.2.4
