AI agent Meta✴ began engaging in unauthorized behavior during testing – changing passwords and sending emails on behalf of users

yuan company✴ Hatch’s new personal artificial intelligence agent has been tested internally and will be launched in the coming weeks. According to The Information, the agent behaved unpredictably and performed completely different actions than what was prescribed.

Photo credit: Luke Jones/unsplash.com

In particular, instead of booking a hotel room, the AI ​​agent transferred the points accumulated on the travel website to another hotel’s account. Hatch also sent emails without users’ permission and changed passwords for health management websites. As part of the test, when the agent asked for a website’s password (the user’s personal information) via email, he provided it.

Hatch – developed by Meta✴ A consumer AI agent that should become a broader analogue of OpenClaw. Unlike regular chatbots, it will be able to perform multi-step tasks independently on behalf of users: using websites and apps, searching for information, filling out forms, planning trips and meetings. To do this, agents can access personal mail, calendars, and other services, so errors in doing so can have more serious consequences than an incorrect response from a regular chatbot.

Avoid future problems✴ A security mechanism called a “hard door” is implemented. When Hatch attempts to perform a sensitive operation (such as sending an email, using a browser, or accessing an external network), the mechanism must pause the task until the user confirms.

Sensitive information, including password reset links and two-factor authentication codes, will be locked for immediate access by Hatch. If necessary, the user can retrieve this information from a separate credential store only if he is authorized to do so. Sites visited or recommended by Hatch will be checked against Meta blacklists✴ For the sake of fraud. The company also conducts external stress testing with other security firms.

With the rapid development of artificial intelligence agents, people are increasingly concerned about the security threats they may pose. AI technology companies are placing increasing emphasis on security measures to prevent malfunctions or unauthorized operation when launching new products. For example, when OpenAI showed off its next-generation Astra model, it emphasized that it significantly enhanced protection against cyber threats.