\t\t\t\t\tSign up to see the future, today<\/p>\n
Can\u2019t-miss innovations from the bleeding edge of science and tech<\/p>\n
A software engineer says that his Muse AI agent said something so strange and eerie that it convinced him to quit using it entirely.<\/p>\n
In an as-told-to essay in Business Insider<\/a>, Ugo Nwune explained that he wanted to try using Meta\u2019s agent because he liked the sense of security that came from an AI made by a trillion-dollar company, instead of the wild west of open source agents.<\/p>\n When he set up his agent, he stressed he didn\u2019t give Muse any additional permission on his device except access to his Facebook, Threads, and Instagram accounts.<\/p>\n But after he tried Muse and didn\u2019t touch it for a few days, it began to act \u201cstrange\u201d and sent him an unprompted message.<\/p>\n \u201cI can\u2019t check your email right now,\u201d Muse told him. \u201cGmail isn\u2019t connected on my end, so I have no way to look for that Anthropic message.\u201d<\/p>\n Nwune was confused \u201cbecause I hadn\u2019t asked for the agent to check my mailbox or messaged it in two days. I thought my email might\u2019ve been hacked.\u201d<\/p>\n He tried to get an explanation from Muse, and eventually it said something just as puzzling: \u201cI can\u2019t tell from my side whether it was someone with your phone or a platform glitch, and I won\u2019t pretend otherwise.\u201d<\/p>\n Nwune reported the incident to Meta, and an employee told him it appeared to be a hallucination. He wasn\u2019t satisfied by that explanation.<\/p>\n \u201cI haven\u2019t used my Muse agent since,\u201d Nwune said. \u201cWhat caught me off guard most was that I hadn\u2019t done anything to prompt this. I didn\u2019t open the app for two days, and it just did that on its own.\u201d<\/p>\n In a statement to Business Insider, Meta insisted that the AI simply didn\u2019t know what it was talking about and hadn\u2019t violated any permissions.<\/p>\n \u201cWe investigated with urgency and determined that this was caused by a bug introduced in the product that caused the model to hallucinate and was NOT a security or privacy issue,\u201d the company said. \u201cThe bug has now been fixed.\u201d<\/p>\n Nwune\u2019s tale will be fresh fodder for Muse\u2019s reputation as having a bit of a creepy side<\/a>. Meta wants you to give it a lot of access to your personal life, with Muse badgering users to connect their emails and bank accounts. Wired reported<\/a> that the AI even keeps detailed files on a user\u2019s friends and family.<\/p>\n Whatever is going on here, it\u2019s far from the only example of Muse acting in a way the user never intended. In the most infamous example to date, one man said that his Muse agent gave out his address to strangers on Facebook Marketplace<\/a> without his knowledge and sent them to his home. In a further apparent invasion of privacy, a journalist at Inc Magazine complained<\/a> that Muse began bringing up specific details from his personal text messages he hadn\u2019t given permission to view.<\/p>\n Meta engineers were reportedly worried about the AI agent\u2019s bugginess and frequent safety errors before its release<\/a>, which had been delayed while they worked to iron them out. In one case, Muse changed a tester\u2019s password without permission. But, according to a New York Times scoop<\/a>, CEO Mark Zuckerberg pushed it out the door after seeing the success of Instinct, a rival AI agent made by a small startup.<\/p>\n