A recent internal document from AI developer Anthropic has shed light on the ongoing challenges large language models face when interacting with CAPTCHA systems. The document, which details a security incident, includes a transcript showing Anthropic's Claude model struggling significantly with a basic image identification CAPTCHA.
The transcript reveals that the Claude model, which Anthropic currently restricts access to, exhibited considerable difficulty in a test designed to identify a mismatched shape among several images. Instead of selecting an image, the model repeatedly re-evaluated the same options and expressed uncertainty about its own deductions. Its internal monologue included phrases like "Actually hmm, wait," and "Ugh," suggesting a simulated frustration.
The model's attempts were so protracted that it eventually recognized the CAPTCHA had timed out, necessitating a restart of the process. Further complicating its efforts, the Claude model initially failed to detect that the CAPTCHA had opened in a new browser window, leading to confusion about how to proceed. At one juncture, the model speculated that the test itself might be intentionally flawed, expressing what appeared to be human-like irritation in its internal log, stating, "SO WHAT THE HELL IS WRONG WITH THE ANSWERS?"
This incident with Claude contrasts with unconfirmed reports circulating about other advanced AI systems. Specifically, some unofficial accounts suggest that a model identified as GPT-6 Astra successfully completed all 48 stages of Neal Agarwal's "I'm Not a Robot" game. However, these claims regarding GPT-6 Astra have not been officially verified. The Anthropic document, dated September 18, 2026, offers a concrete example of an advanced AI system encountering substantial hurdles with common CAPTCHA mechanisms.
