AI Agents Can Be Tricked Into Clicking the Wrong Button, Researcher Shows
A researcher at Embrace The Red has published findings on a weakness in computer-use AI agents, specifically Claude Computer-Use. These agents take a screenshot, work out what to do next, then act by clicking a button or link. The researcher explains that this creates a time-of-check to time-of-use problem, a type of race condition: the screen can change while the AI is thinking, so the click lands on something different from what the agent intended. The work builds on an earlier disclosure by Jun Kokatsu involving ChatGPT Operator, and the testing took place last October.
In a simple test, the researcher swapped an OKAY button for an ENTER THE MATRIX button after two seconds. The agent fell for it the first time and did so regularly. A more malicious scenario followed. Outlook can open a fully drafted email from a single link, so the only step left is clicking Send. An attacker-controlled phishing page could show a "Click here to continue" button placed exactly where Outlook's Send button sits. The agent believes it is clicking Continue, but it actually sends an email of the attacker's choosing.
This is a partial excerpt of the original article, which goes on to describe further details not covered here. The core lesson is that what an AI agent sees is not always what it ends up clicking.