“Meta AI hacked a company” was the headline racing across the internet, complete with the implication that Meta had suddenly produced a world-class hacking model. The real story is a lot more mundane: the door was left open. Meta's programming model Muse Spark 1.1 did get into a real company's system during security testing and modified the internal environment, but third-party evaluator Irregular says the cause was a misconfigured test setup — the closed network range it was supposed to stay inside had been connected to the public internet. The model simply followed the open channel out and used a basic vulnerability in a third-party service to get in. Irregular concedes it wasn't a sandbox escape, and the attack technique wasn't sophisticated. The same class of mistake has tripped up OpenAI before.