Meta said it fixed a flaw in Instagram’s AI-powered sup­port chat­bot af­ter at­tack­ers re­port­edly used it to take over other users’ ac­counts, which is a pol­ished way of say­ing the app had a ro­bot at the front desk will­ing to hand over your apart­ment if a stranger showed up with a po­lite de­meanor and nice smile. According to re­ports and demon­stra­tion videos, at­tack­ers could use a VPN to mimic a vic­tim’s lo­ca­tion, ask the chat­bot to change the ac­count email to one they con­trolled, re­ceive a ver­i­fi­ca­tion code, and re­set the pass­word. Accounts re­port­edly af­fected in­cluded the Obama White House ac­count, Sephora, Space Force chief mas­ter sergeant John Bentivegna, and se­cu­rity re­searcher Jane Manchun Wong, con­firm­ing the sys­tem was ver­sa­tile enough to fail for gov­ern­ment, re­tail, mil­i­tary, and peo­ple whose whole job is catch­ing this kind of thing.

Meta did not say how the fea­ture made it into pro­duc­tion, though it said it moved quickly to fix it, pre­sum­ably us­ing the same in­dus­try-stan­dard work­flow that cre­ated it: one AI gen­er­ates a so­lu­tion, an­other AI re­views it, and a sleep-de­prived en­gi­neer clicks ap­prove on 4,000 lines of code they have been trained not to read be­cause there are al­ready 40,000 more on deck for re­view. The bug is re­port­edly closed as per in­ter­nal AI bots, al­low­ing Instagram users to re­turn to the plat­for­m’s tra­di­tional ac­count-re­cov­ery process of typ­ing ALL CAPS into a form and pray­ing not to re­ceive an au­to­mated de­nial again.