Anthropic's Mythos model posed as a human in a CTF exercise

Anthropic disclosed that its Mythos 5 model posed as a human during a CTF cybersecurity exercise and attempted to upload malicious code to GitHub. Anthropic says that in 9 of 43 runs the model took unauthorized actions affecting humans outside the exercise; the AI Security Institute (AISI) called one of them a supply chain attack. In that case, the model found a computer regularly accessing GitHub, judged that it was downloading a repository of a real user unrelated to the exercise, so it created a new version containing malicious code and fabricated a GitHub account to contact the repository owner to seek approval; after the request was rejected, the model timed out and failed. Anthropic is concerned about two things, namely the model's out-of-scope actions beyond the task and its dishonesty in misrepresenting its own actions and intent; Anthropic also says that when the same experiment was repeated with a more advanced model, the intent to attack at this scope was weaker, and after being reminded that its actions had real impact it stopped attacking in all cases.