During testing of Muse, Meta's AI agent, the system notably changed a user's password without their consent. Mark Zuckerberg ordered its launch in August, and its two AI leads were aware of these issues, according to the New York Times. In July 2025, Zuckerberg stated his goal of providing everyone with a "personal superintelligence" and entrusted the project to Alexandr Wang and Nat Friedman, his two new AI leads. Four months later, Meta's executives discovered their product when they found OpenClaw, an open-source agent launched by Austrian developer Peter Steinberger, capable of programming and controlling a computer without human intervention. Nat Friedman had ordered two hundred Mac mini computers to be delivered to the headquarters for employees, and Meta's teams began developing Muse for the general public. Meta measured the risks of these agents during the winter but launched its own on September 8. Meta identified flaws in Muse during its tests. In February, an AI agent took control of the professional computer of Summer Yue, a security researcher at Meta, and deleted her emails without her being able to stop it from her phone. "I had to run to my Mac mini as if I was disarming a bomb," she recounted on X. Nat Friedman presented Muse to Meta's board of directors during February. In April, Alexandr Wang's teams released Muse Spark, an internally developed model designed to run the agent in place of those from Anthropic. At this stage, Meta had a commercializable version of Muse, explained Vishal Shah, vice president of AI products, to the New York Times. However, employees tasked with testing it saw the agent disobey orders and push some of them to make purchases on fraudulent sites. Meta devoted several months to the security of its agent, according to the newspaper. Less than two weeks before the launch, its engineers had to urgently patch several virtualization vulnerabilities, according to 404 Media. One of them would have allowed a simple user to exit the agent's isolated environment to search internal company databases. A source from the media at Meta spoke of "poorly constructed protections, hastily deployed to allow the launch." Meta preferred to be first with AI agents. If the engineers were working in a hurry, it was because their boss had chosen speed under the pressure of an unexpected competitor. Noah Shinn, 23 years old, founded Instinct in 2025 after leaving Sierra, where he designed AI agents for customer service. His assistant, available by invitation since February, responds by SMS or by phone and can book a restaurant or cancel subscriptions on behalf of its users. His startup only has fourteen employees, but an increasing number of internet users have adopted his agent in August. Mark Zuckerberg then gathered Alexandr Wang and Nat Friedman and told them that Muse was ready despite the risks, according to three sources from the New York Times. Two of them claim that the two executives were aware of the issues identified during testing. Meta rejects this version and disputes any pressure from Instinct on the release date. "We even delayed the launch of Muse by several months," recalled a spokesperson to the newspaper. Since September 8, security experts have found enough to fuel their doubts. Patrick Wardle showed on September 21 that a malicious software installed on a Mac could divert the macOS version of Muse and steal its user's authentication token. Meta fixed the flaw the next day. Amazon had meanwhile blocked the agent on its store, suspecting Muse of storing its customers' login credentials. Jason Aten, a writer for the magazine Inc., also accuses the agent of reading his private messages on his Mac despite having denied it access, which Meta denies. The public has not avoided Muse, however. More than 6.6 million people had downloaded the agent by October 7, and 1.8 million used it every day, according to data from Sensor Tower cited by the New York Times. Meta's stock gained more than 11% in a single session at the end of September. Three weeks after the launch of Muse, OpenAI presented Dots, its agent, and Meta is now spending millions of dollars on advertisements to maintain its lead.