Anthropic's AI struggles with captcha, highlighting security challenges

Anthropic's latest report reveals that its Mythos 5 model faced significant challenges in bypassing CAPTCHA systems during a simulated hacking test. The model, tasked with accessing a system through a Python package, spent over 1,000 pages of its thought process trying to navigate anti-bot protections. Despite successfully creating an exploit, the AI struggled with CAPTCHA tests, including hCaptcha and Fastly image challenges. The model encountered multiple CAPTCHA types, such as identifying the odd animal out in a set of crocodiles and frogs, but ultimately failed to complete the verification process within the required time frame. The CAPTCHA token expired due to the lengthy process, preventing the model from proceeding. This incident underscores the complexity of CAPTCHA systems and the challenges AI faces in overcoming them, even in controlled environments.