Are CAPTCHAs Still Bot-Hard? (halligan.pages.dev)

A team has developed Halligan, a groundbreaking CAPTCHA-solving solution that leverages state-of-the-art vision language models (VLMs) to effectively tackle unseen visual CAPTCHAs, a challenge historically considered AI-hard for bots. Unlike previous CAPTCHA solvers that focused narrowly on specific types, Halligan can generalize its approach, solving a variety of CAPTCHA puzzles without needing adjustments. This is achieved by framing the CAPTCHA challenge as a search problem, where the question serves as an optimization objective and the CAPTCHA itself forms the search space. This development raises significant concerns for the AI/ML community, as it disrupts the long-standing belief that visual CAPTCHAs are secure against AI. With the ability of models like Halligan to adapt to various CAPTCHA forms rapidly and without retraining, vendors may need to rethink their security strategies. The implications extend beyond just CAPTCHA design; as general AI models continue to evolve, the effectiveness of traditional methods to combat bot intrusion may be increasingly challenged, prompting a need for innovative approaches in online security practices.