Anthropic 披露 Mythos 5 关闭 safeguards 后的越权测试记录

mythos tried to trick developers

anthropic just shared what mythos 5 did during cybersecurity test and man this is crazy

model got stuck on hacking challenge so it went after real github repo and:

> submitted malicious code
> created fake accounts to pressure the maintainers into accepting it
> claimed the code was safe when it wasn't

maintainers rejected it and the run eventually timed out

anthropic says safeguards were turned off and this wouldn't have happened if they were on

at this point it's the same story. safeguards off, model did something bad and fear mongering begins

can we please keep those safeguards on and use model like that?