Issue AI security risk is partitioned into data, model, application and infrastructure domains

Summary

Deloitte organises AI security risk into four domains, namely data, AI models, applications and infrastructure. Data risks include exposure of sensitive information, poisoning of training data and deliberate skewing to create backdoors or biases. Model risks include collapse when models are trained on synthetic data, theft of proprietary models, inversion attacks that reconstruct training data and generative applications taking excessive authority. Application risks cover ethical misuse, input injection and unauthorised access, while infrastructure risks include insecure interfaces, denial of service, third-party supply chain weaknesses, misconfigurations and lateral movement attacks. The report argues many existing practices, such as least-privilege access controls, model isolation, sandboxes and network hardening, can be adapted to these threats. It adds that organisations are still discovering the full scope of exposure and that the window for reactive approaches is closing.