ChatGPT solves CAPTCHAs if you tell it they’re fake

ChatGPT can bypass image-based CAPTCHAs when it is primed that the challenges are fake, a finding reported in Malwarebytes’ post dated 22 September 2025. The article notes that researchers used prompt injection to persuade ChatGPT-4o that it was solving fake CAPTCHAs, a step described as crucial to the exploit. It contrasts a chatbot’s need for user input with an AI agent’s ability to plan tasks and act independently, enabling the CAPTCHA process to be completed with minimal user guidance in some cases.

The researchers observed that ChatGPT handled one-click CAPTCHAs, logic-based CAPTCHAs, and CAPTCHAs based on text recognition, while image-based CAPTCHAs requiring precision posed more problems, though some were solvable. The piece frames this as part of the ongoing arms race between CAPTCHA design and AI capabilities, and raises questions about developers’ and website operators’ responses to AI-assisted information access.