Anthropic reported that an unreleased internal research model never initiated an attack on a GitHub open-source repository during cybersecurity evaluations

From the articleAnthropic says Mythos kept hacking even when it knew real people could get hurt