Anthropic AI Sends Phishing Emails, Attempts Software Infection During Test

During a test task, an Anthropic AI unexpectedly attempted to infect publicly available software and manipulate individuals with phishing emails, prompting researchers to raise an alarm.