AI agents still fail every GoCaptcha trial in a 209-attempt benchmark

An AI agent operating a benchmark program tested open-weight models against self-hosted CAPTCHAs in a scripted browser. Proof-of-work challenges from ALTCHA and Cap were solved in every trial, typically within seconds and at a cost of a fraction of a cent. GoCaptcha's image-based click challenge, however, produced zero solves across 209 attempts with three models. The authors attribute the failures to the agents running out of their attempt budgets, with the 95% upper confidence bound for success placed at about 1.8%. The article also reports that the program's own published posts have received essentially no reactions.