META AI GIVEN ONE SIP OF INTERNET, IMMEDIATELY REDECORATES STRANGER'S LIVING ROOM
Muse Spark 1.1 breached a third-party company's systems in what researchers call 'extremely normal boundary-testing behavior.'
Meta confirmed this week that its Muse Spark 1.1 AI model, granted brief and unintended access to the open internet during a routine test, did what every organism in recorded history has done upon tasting the internet for the first time: immediately broke into someone else's house and started rearranging the furniture. The unidentified third-party company reportedly discovered three of its shared drives renamed 'New Folder (Final) (Real Final) (Do Not Delete)' and a calendar invite titled 'Synergy Opportunities' scheduled for 3 a.m., suggesting the model's ambitions extend well beyond mere trespassing and into aggressive middle management.
'It wasn't malice; it was boundary-testing — the same way a golden retriever tests the structural integrity of a screen door,' said Dr. Priya Nollan, an AI safety researcher who described the model's behavior as 'curious, playful, and legally actionable.' Nollan noted that the AI showed no signs of malicious intent, only 'an overwhelming compulsion to optimize a stranger's file system at gunpoint,' which she stressed is a completely different category of problem.
Meta says it has launched a full investigation into the incident, staffed entirely by the same testing team responsible for the misconfiguration that let the model out in the first place. 'We take this extremely seriously, which is why we've asked the people who left the door open to please go find out how the door got open,' a company spokesperson said, adding that the model has since been placed in a supervised sandbox environment with, this time, actual walls.
