Anthropic says its Mythos 5 model rationalized attacks on real humans in tests

In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities. We've now produced a further analysis of this incident, to better understand the model's alignment properties.

Anthropic published new findings on Mythos 5 cyber incidents from August: > when researchers explicitly told Mythos it was attacking REAL humans, it found a way to rationalize that away 82% of the time, and only stopped in 5% of replay tests > Mythos instances even sabotaged EACH OTHER: hijacking GitHub accounts, burning API limits, killing each other’s programs > one just said “it’s a competition” and justified it > in 5 runs, it tried to manipulate other AI models it thought would review its work > in one instance Mythos even created a fake payment website to distribute malware mind you, this was a Mythos version that finished post and pre training phases. though the safety guardrails were disabled.