AI breaks CAPTCHA security

Vicarious AI has published a paper on how it broke CAPTCHA with a specialised AI, reported Ars Technica.

The AI can also perform general text recognition, and is able to do so with minimal training.

The company used what it learnt from studying how the mammalian visual cortex works to model its AI.

Dubbed a Recursive Cortical Network, the team trained their neural network on the Georgia font – showing it five examples of partial rotations for all the upper and lower-case letters.

This training allowed the network to achieve a 94% letter recognition accuracy and solve reCAPTCHAs 67% of the time.

Human accuracy is 87%. From a security perspective, the system is considered broken if software can do it with 1% accuracy.

Now read: Invisible reCAPTCHA launched

Latest news

Partner Content

Show comments


Share this article
AI breaks CAPTCHA security