To test whether this idea actually works, The company evaluated five leading AI models: Opus 4.8, Gemini 3.1 Pro, GLM 5.2, DeepSeek V4 Pro and Kimi K2.6. They conducted 152 attack simulations inside a custom-built AWS cloud environment containing nearly 300 resources, including services like S3, EC2, Lambda, DynamoDB and Secrets Manager.