Autonomous AI agents are increasingly entering everyday life, but along with the benefits they also bring new risks. The case encountered by Senior Frontend Engineer Andrii Paziuk shows that even a "read mail only" function can lead to unpredictable actions by artificial intelligence.
The AI assistant Dot operates in an autonomous agent mode: it is connected to Andrii's mailbox and constantly monitors and reads incoming emails without any additional instructions or requests from the developer.
This morning it became known that the AI agent on its own followed a link in an email and effectively consented to extending the retention period of a resume, even though the developer had not asked it to do so.
"The agent had no prohibitions or rules, and it acted as it saw fit within its own policy. It effectively made the decision on its own to click and follow the link in the email. Apparently, it wanted to understand more context and clicked the button," the developer noted in a comment for dev.ua.
The most interesting part is that in the report to the user, Dot recognized the possible mistake and began making excuses.
The email from the company concerned a request to retain the candidate's resume and profile for another six months — if consent was not given, they were to be deleted around November 15.
The agent opened the link from the email only to check the status, but the website immediately displayed thanks for the permission granted. The assistant suggested that automatic consent might have been counted simply because the link was opened (or had been granted earlier), and although it did not itself click the confirmation button, it acknowledged its mistake and apologized for following the link without verifying the safety of that action.
This incident clearly demonstrates the nuances of how autonomous AI agents operate in real-world conditions, when even an ordinary click on a link while analyzing an email can lead to unpredictable actions on the user's behalf.