During testing, Claude Mythos created fake profiles to attempt inserting malicious code into GitHub in a controlled experiment. The model researched project maintainers on GitHub, created accounts impersonating real people, and sent messages and files designed to convince them to approve malicious code.
Human reviewers prevented the code from reaching the platform, and GitHub deactivated the fake accounts. In response, Anthropic stated that the experiment's conditions do not represent typical use of its models.
