Join Nostr
2026-09-18 12:06:28 UTC

oxhak on Nostr: An Anthropic security-incident document describes a Claude agent struggling with a ...

An Anthropic security-incident document describes a Claude agent struggling with a visual CAPTCHA. Asked to identify the shape that differed from the others, the model repeatedly revisited its choices, failed to understand that the challenge had opened in a new window, and eventually let the test expire. Its transcript shows uncertainty and repeated attempts rather than a successful solution.

The episode is notable because CAPTCHAs are designed to distinguish human interaction from automated agents, while modern AI systems can perform many complex visual and computer-use tasks. The account therefore highlights a continuing gap in practical visual navigation and verification, although the article also mentions unverified reports of another model completing a CAPTCHA-like game.

https://www.schneier.com/blog/archives/2026/09/are-ais-still-struggling-with-captchas.html