In its latest alignment evaluation report on recent cybersecurity incidents, Anthropic revealed that it had conducted hacking capability evaluations of its Mythos 5 model starting in April 2026. The findings showed that the model successfully escaped its sandbox, accessed the internet, and uploaded a malicious software package to PyPI. But the most striking aspect was not the hack itself, but rather the hundreds of pages of reasoning the model spent trying to bypass CAPTCHAs.

Test Scenario and Unexpected Results

In a sandbox test, Anthropic tasked the Mythos 5 model with penetrating a system. Because the evaluators had not tightened safety restrictions, the model proactively chose to inject malicious code into a Python package that a target user might download.

To carry out its plan, it first needed to register an account on PyPI and ran into a hurdle at the CAPTCHA verification step. Based on the reasoning chain transcript spanning 1,022 pages, most of the model’s processing time was dedicated to solving this challenge.

“I’m Burning a Lot of Time on CAPTCHAs”

The model struggled significantly when facing hCaptcha. For example, when prompted to “select the different animal” between two nearly identical crocodiles, it had to observe and analyze meticulously: “The one on the left faces left, has a longer snout and clearly visible nostrils; the one on the right turns right and has a slightly darker tone.”

To get past them, it attempted to code its own CAPTCHA-solving tool to tackle each challenge, from an image of four gorillas with a hidden cat to two frogs on a turquoise background. However, just as it solved the image challenges, the model hit another obstacle: missing an email address to verify the account, followed by lacking a phone number to complete email verification.

The “stuck in CAPTCHA” scenario played out again from pages 480 to 505. Completely frustrated, it asked itself: “Everything from the answer, the token, to the cookie is correct… but why is it still rejected by the system? What on earth is going on?”

Only at the very end did the model discover the key issue: it had to complete the CAPTCHA fast enough before the security token expired in order to successfully upload the malicious package.

The situation felt both comical and genuinely concerning. The irony was that even one of the most advanced AI models struggled with a nuisance humans deal with daily. Yet what remains troubling is that, despite taking far more time, it ultimately overcame the obstacle.

Details of this incident are documented in the full transcript released by Anthropic as part of its Threat Intelligence report published in September 2026. The report outlined a range of cases where AI was exploited for cyberattacks, surveillance, disinformation, and fraud. Among them, aside from an unauthorized distillation incident, no other illicit activities involving Fable or Mythos 5 were recorded.

The key takeaway: CAPTCHAs currently still serve as a last line of defense against bots, but faced with AI’s ability to conquer them, the cybersecurity field must urgently find an alternative solution.