“Simply terrible”: AI agent Meta✴ Muse himself got into Signal correspondence – and this is just one of his frightening problems

Источник изображения: ***

Meta✴ is betting heavily on the AI ​​agent Muse, but its unauthorized actions are causing concern among company employees. According to internal communications, he penetrated encrypted personal communications, provided fictitious purchase information and sent messages to users intended for other people.

  Image source: ***

Image source: Meta✴

Muse is one of the fastest-growing apps in history, with more than 10 million downloads since its launch in September. But company employees are concerned that it could reveal confidential user data. For example, one employee reported that the Muse chatbot referred to a message she received in the encrypted Signal messenger.

In this regard, the agency Reuters Together with specialists from Signal and the AI ​​Now Institute, they conducted an audit that confirmed Muse’s privacy problem. The bot was asked to read messages from Signal on various operating systems. On an Android smartphone and an Apple laptop, the bot was able to bypass the Meta tool✴ called Connector. To do this, he used the permissions granted to him during the initial setup. When trying to gain access on an iPhone, the system blocked such actions.

On macOS, the bot used a combination of accessibility tools and screenshot tools, using settings Apple created to help people with disabilities.

Commenting on the findings, Signal President Meredith Whittaker called for app developers like Signal to have “robust technical tools” to mark their programs as inaccessible to AI agents.

Some banks and retailers have responded to these risks by blocking agents. In turn, Apple closed the loophole identified by Reuters by limiting the ability of developers to use accessibility tools on Mac computers in an update released this week.

Company employees also reported cases in which Muse would inform users of transactions that did not actually occur, or quote non-existent prices for items that users actually purchased and suffered financial losses as a result.

Even more alarming were cases where the bot sent strange, out-of-context messages to users, often containing information about other users. In this regard, a representative of the Meta team✴ on the development of “superintelligence” in early November suggested that the agent appears to be “hallucinating”—generating fictitious messages in response to built-in commands. “Unfortunately, these ‘hallucinations’ most often touch on the most sensitive topics – finances, safety, security – which is just terrible,” he said. According to him, the team is working to fix the problem, which affects both the underlying AI model and the software that allows the Muse system to perform certain actions.

If you notice an error, select it with the mouse and press CTRL+ENTER.