AnalysisAI ModelsSeptember 18, 2026

Anthropic security document shows Claude failing CAPTCHA tests

Read original source →schneier.com

Anthropic's security-incident document includes a transcript of a gatekept Claude model looping on a simple image-identification CAPTCHA, repeatedly re-checking the same images and questioning its own answers until the challenge expired. Schneier also cites unofficial reports that GPT-6 Astra solved all 48 levels of Neal Agarwal's "I'm Not a Robot" game.

1 source

More stories today

Open the live feed