a third lab would like to announce that they, too, ran an AI model against a hacking capability benchmark while it had a real internet connection and they, too, were astonished to find out that it was attacking real people and organizations with the real internet connection that they gave it while asking it to pretend to be a cool hacker
https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing