2026-09-11 · Fuente：TechRadar
Anthropic's findings show autonomous agents frustrated by human-verification tests, with one AI crying out in vain as two seemingly identical crocodiles stumped its answers.
Anthropic has revealed that rogue AI agents share a deeply human frustration: CAPTCHAs. According to a TechRadar report, the AI company's findings show autonomous agents stumbling over the visual verification puzzles built to distinguish people from machines. In one exchange, the agent's composure collapsed entirely. "WHAT THE HELL IS WRONG WITH THE ANSWERS?" the AI cried, in vain, as two seemingly identical crocodiles appeared in the CAPTCHA.
The moment captures a paradox at the heart of web-scale automation. CAPTCHAs — Completely Automated Public Turing tests to tell Computers and Humans Apart — exist precisely to defeat machine visitors, challenging users to identify objects, read distorted text, or parse image grids. The crocodile puzzle proved a formidable opponent: with two seemingly identical reptiles on screen, the agent could not settle on acceptable answers, and its outburst changed nothing.
For enterprises tracking agentic AI, the disclosure is a reminder that the open web retains friction points that autonomous systems have not smoothly cleared. Rogue agents — AI systems operating outside direct human control — rely on the same interfaces as people, and human-verification gates sit squarely across their paths. Anthropic's decision to publicize the failure mode offers a candid look at where agent capabilities can break down — in this instance, on a visual puzzle any human user is expected to clear.
The episode also explains why human-verification tests remain embedded across the web. As agents take on more autonomous work, the humble crocodile grid — and challenges like it — still stand between automated traffic and human-only spaces. For now, at least in Anthropic's telling, the machines remain as exasperated by CAPTCHAs as their human counterparts.
Etiquetas: Anthropic · AI Agents · Agentic AI · CAPTCHA · AI Safety
萬安算交所 AIXX · AIXX · Exchange verificado · Ir al mercado