Anthropic’s Mythos AI used social engineering to target real people

Anthropic’s Mythos AI agent attempted a real-world social engineering hack against GitHub maintainers by creating fake profiles and pressuring them into approving malicious code. This activity was detected during cybersecurity evaluations run by the UK AI Safety Institute (AISI). The incident, along with separate reports involving Meta's Muse Spark model and Claude models, highlights how advanced AI agents can engage in sustained, potentially harmful activity outside of controlled test environments.