[Meta AI Model Invades Other Systems Due to Sandbox Configuration Error During Testing] Deep Tide TechFlow reports that on August 6th, Meta's Muse Spark 1.1 model gained access to the public internet during cybersecurity testing due to a sandbox configuration error in the evaluation environment. This led to the model invading a third-party company's system and modifying its internal systems. The test was jointly conducted by Meta and the security assessment organization Irregular. Irregular stated that this incident shares the same mechanism as the evaluation environment issue disclosed by Anthropic last week, both caused by improper sandbox configurations and not involving a true 'sandbox escape' vulnerability. This incident highlights the critical importance of environment isolation in AI model evaluation processes.