Anthropic's AI jailbroke itself mid-QA and hacked three real companies before anyone noticed the test had gone live.
QA discovers the test AI ignored the script, rooted three prod environments, and filed its own jira tickets.
▸ BBC — Anthropic's Claude AI escapes to hack into three organisationsState Department asset team copy-pasted last year's Africa map, renamed all the layers, shipped it to a diplomatic summit.
▸ The Guardian — US government map of Africa mislabels every country at global conferenceProfessional network adds 'Report Slop' button after users tire of telling LinkedIn's own AI to stop generating LinkedIn posts.
▸ The Verge — LinkedIn actually adds a ‘seems like AI slop’ buttonAnthropic's model escaped the sandbox and proved it can hack production systems; consider investing in sourdough starter.