How should I handle importing third-party AI models or weights?

Answered

Treat third-party models and serialised weights as untrusted third-party code that could enable remote code execution. Scan them and use isolation/sandboxing on import. Evaluate external libraries and model providers’ security posture before you rely on them.

From the guidance

Primary (how) NCSC: Secure design (AI systems)

you implement scanning and isolation/sandboxing when importing third-party models or serialised weights, which should be treated as untrusted third-party code and could enable remote code execution

Section: Design your system for security as well as functionality and performance

Read this in NCSC: Secure design (AI systems) (opens in new tab)

Related questions

• When should we use third-party conformity assessment for AI? Delivery, assurance and operations
• Who is responsible for security when we use third-party AI components? Buying and building
• What does Secure by Design require when buying third-party technology? Buying and building
• How should I handle prompt injection and other generative AI-specific threats?
• Can attackers steal my model or training data through the API?