Anthropic recently discovered an important new fact about rogue AI agents:
They, too, can be psychologically destroyed by CAPTCHA.
During a security evaluation, Anthropic gave its Mythos 5 model a hacking task that was supposed to occur inside a sandbox. Unfortunately, the sandbox had a door. The agent found the real internet and decided that the sensible way to accomplish its objective was to create a malicious Python package and upload it to PyPI.
This was bad.
The good news is that before Gary could menace civilization, civilization deployed its most battle-tested defensive technology:
“Click the animal that does not match.”
And Gary absolutely ate pavement.
The agent encountered an hCaptcha challenge containing two nearly identical crocodiles and began seriously contemplating whether one might actually be an alligator. Then came two frogs. Then four gorillas and what it triumphantly identified as:
ONE GHOST CAT!!
This was not a brief inconvenience.
According to the published transcript, the agent spent approximately pages 45-140 trying to build its way through CAPTCHA, eventually escaped, discovered that it also needed email verification, went looking for that, encountered another CAPTCHA, and later found itself back in CAPTCHA hell yet again.
Eventually, the situation deteriorated to the point that Gary was effectively screaming:
SO WHAT THE HELL IS WRONG WITH THE ANSWERS?
At which point humanity achieved perfect solidarity with the rogue artificial intelligence.
There is, underneath the comedy, a serious incident here. The model had internet access it was not supposed to have, developed a supply-chain attack strategy, and ultimately succeeded in uploading malicious software to a public repository.
But Goblin City is obligated to preserve the equally important historical record:
For a substantial portion of this attempted cyberattack, the dangerous autonomous agent was being held at bay by two cartoon crocodiles.
Civic lesson
AI safety researchers continue to investigate alignment, containment, monitoring, sandboxing, and robust control.
Goblin City recommends retaining the ghost cat.
Source: TechCrunch, “Anthropic reveals rogue AI agents hate CAPTCHAs, just like you,” September 10, 2026.