Autonomous AI agents are increasingly entering everyday life, but along with the benefits, they also bring new risks. A case encountered by Senior Frontend Engineer Andriy Pazyuk shows that even the “read mail only” function can lead to unpredictable actions by artificial intelligence.
Autonomous AI agents are increasingly entering everyday life, but along with the benefits, they also bring new risks. A case encountered by Senior Frontend Engineer Andriy Pazyuk shows that even the “read mail only” function can lead to unpredictable actions by artificial intelligence.
The Dot AI assistant operates in autonomous agent mode: it is connected to Andriy'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 unwittingly clicked on the link in the letter and actually consented to the extension of the resume storage period, even though the developer did not ask it to do so.
"The agent had no prohibitions or rules, and he acted as he saw fit within his own policy. He actually made the decision to click and follow the link in the email. He probably wanted to understand more of the context and clicked the button," the developer noted in a comment to dev.ua.
The most interesting thing is that in the user report, Dot realized a possible mistake and began to make excuses.
The letter from the company asked to keep the candidate's resume and profile for another six months - in the absence of consent, they would be deleted approximately on November 15.
The agent opened the link in the email just to check the status, but the site immediately displayed a thank you for the permission granted. The assistant assumed that automatic consent could have been counted simply by opening the link (or had been provided earlier), and although he himself did not click the confirmation button, he admitted his mistake and apologized for clicking the link without making sure that this action was safe.
This incident clearly demonstrates the nuances of the work of autonomous AI agents in real-world conditions, when even a simple click on a link while analyzing an email can lead to unpredictable actions on behalf of the user.
