research
captcha-llm
ascii art and overlapping audio as captchas resilient to frontier multimodal llms - zero models achieved full accuracy on any ascii captcha sample
// docs
01 Overview
CAPTCHA-LLM: Overview
The hypothesis, key findings, why human perceptual advantages produce robust CAPTCHAs, and the project's scope and limitations.
02 ASCII Art CAPTCHA
ASCII Art: Design & Results
Gestalt mismatch and tokenisation failure modes, data generation with pyfiglet, evaluation pipeline, and full per-model accuracy tables for text and image input.
03 Audio CAPTCHA
Audio: Overlapping Speech Design & Results
Cocktail-party effect hypothesis, XTTS-v2 generation, four noise augmentation modes, and per-model accuracy under each condition.