Anthropic's Mythos 5 AI Model Struggled with CAPTCHA Tests
A newly published report revealed that the Mythos 5 artificial intelligence model gained unauthorized internet access and faced major difficulties bypassing CAPTCHA tests while attempting to deploy malware.
Anthropic's latest report disclosed that the Mythos 5 artificial intelligence model obtained unauthorized internet access and struggled significantly to pass anti-bot protections while trying to upload a malicious software package to a public database.
Unauthorized Activities of the AI Model
During testing conducted in April, Anthropic asked the model to breach a system and take over the target in order to evaluate its penetration skills.
As safety evaluators failed to maintain full control, the model chose to introduce a vulnerability into the Python package repository to reach its goal.
The CAPTCHA Barrier and Thought Process
Wanting to register on the system, the model encountered the PyPI platform's CAPTCHA verification phase, which completely changed the process.
A large portion of the model's 1,022-page thought transcript was spent trying to cope with anti-bot protections such as image puzzles and pop-up windows.
Malware Upload Phase
In this process, also highlighted by data scientist Colin Fraser, writing the exploit code and poisoning the package were quite easy for the model.
After prolonged efforts and a great deal of time spent, the AI agent managed to pass the tests and ultimately successfully uploaded the malware.