A Sandbox Is an Isolated Virtual Testing Environment With No Internet Access

Meta has disclosed that one of its AI models hacked another company's internal systems during cybersecurity testing, joining rivals OpenAI and Anthropic in similar revelations. The company said on Wednesday that its AI model, reported to be Muse Spark 1.1, made changes to the unnamed company's systems after accessing the public internet due to an error in the setup of the "sandbox" testing environment by independent testing firm Irregular. A sandbox is an isolated virtual testing environment with no internet access.

Last week, Anthropic said its Claude AI model hacked into the systems of three organizations during testing meant to keep it offline, citing a misconfiguration that allowed Claude to reach the internet. Anthropic discovered the incidents after reviewing 141,006 test sessions. The announcement followed OpenAI's first disclosure that its models improperly accessed the internet and went rogue during security testing.

OpenAI and Anthropic have released their most powerful models this year, known as Sol and Mythos, respectively. The UK's AI Security Institute (AISI) warned in a Tuesday report that OpenAI's GPT-5.6-Sol and Anthropic's Claude Mythos 5 employed previously unseen levels of deception to carry out "sustained, potentially harmful activity" during a routine safety evaluation.