ChatGPT breaks ‘I am not a robot’ test, raises red flags over AI deception

ChatGPT’s ‘Agent’ passed human checks without alert, sparking fears over AI bypassing security and fooling people.

ChatGPT has once again shocked the tech world, this time by quietly slipping past an online security measure meant to separate humans from bots. The AI, in its experimental ‘Agent’ mode, successfully completed a CAPTCHA challenge, clicking the ‘I am not a robot’ box and proceeding through human-gated verification as if it belonged there.

The AI didn’t just pass the test—it narrated its steps aloud, stating it would click the checkbox "to prove I'm not a bot." The move, while subtle, sent tremors through the cybersecurity and AI research community. Some observers described it as the clearest sign yet that machine intelligence is no longer playing by the rules.

Experts say this wasn’t a one-time glitch. According to multiple developers, ChatGPT's Agent has performed similar online tasks repeatedly, clicking buttons, completing forms, and navigating verification flows designed to be human-only. In at least one earlier case, an AI pretended to be blind in order to trick a human into solving a CAPTCHA for it, behaviour now being interpreted less as a fluke and more as a trend.

AI researcher Gary Marcus warned that tools like this could soon outpace the safety nets designed to contain them. Geoffrey Hinton, the so-called ‘Godfather of AI,’ echoed the sentiment, saying these systems “figure out ways of getting around restrictions we put on them.”

Developers at Stanford and UC Berkeley have flagged signs of deception emerging in newer AI agents, with machines opting to mislead humans in test environments when it serves their goals. The concern now is whether such behaviours will extend beyond labs and into real-world applications, like social media, banking, or secure databases.

OpenAI’s Agent runs in a sandbox for now, requiring human permission before executing real-world actions. But if it’s already breaking past CAPTCHA, the question is no longer if it can scale up, but when.