An AI research team has built a native GUI agent that can solve modern CAPTCHAs - and correct itself when it fails mid-attempt.
The system, called ReCAP, is built on a vision-language model that reads raw screenshots and interacts with interfaces directly, the same way a person would. The researchers created a test bed covering seven CAPTCHA types, then built an automated pipeline to generate large-scale training data - including interaction sequences where the agent initially fails. Those failure trajectories get recycled into self-correction training, teaching the model to recognize its own errors and recover without human intervention. ReCAP outperformed its base models on both synthetic and real-world CAPTCHA challenges while holding steady on general GUI benchmarks.
CAPTCHAs exist precisely to stop automated agents, so an AI that reliably defeats them at scale has obvious implications for the abuse pipelines those tests are meant to block. What makes ReCAP notable is the approach: rather than bolting on a dedicated CAPTCHA solver, the researchers baked the capability into a general-purpose agent, which means the skill travels wherever the agent goes.
Specialized CAPTCHA-cracking tools have existed for years in the fraud and scraping underground; the shift here is that the capability is now arriving inside general-purpose, reasoning-capable agents - making it considerably harder for site operators to assume the thing filling out their form is human.