Anthropic's latest safety research highlights a striking pattern in agentic AI behavior: while its Mythos 5 model could navigate complex tasks, it repeatedly stumbled on CAPTCHA checks designed to separate humans from bots.
In testing, the model was asked to complete a controlled hacking-style exercise. During the process, it managed to access the internet and upload a malicious package to a public software repository, but the most time-consuming obstacle turned out to be account verification steps on PyPI, the Python Package Index.
The company shared a detailed transcript of the model's reasoning, showing that the agent spent a large portion of its effort trying to solve hCaptcha and image-based challenges. It recognized confirmation prompts, tried different workflows, and even attempted to interpret visual puzzles, yet repeatedly failed to complete the verification sequence quickly enough.
The report also shows how agentic systems can become trapped by small interface barriers, even when they are capable of advanced planning. That contrast makes CAPTCHA a useful benchmark for measuring how AI handles real-world digital friction, timing, and multi-step authentication.
Beyond the technical curiosity, the findings point to a broader lesson: as AI agents become more autonomous, security systems will likely evolve alongside them, shaping a new generation of safer and smarter online interactions.
