Digest · past 24h

1. Anthropic and OpenAI agents attacked real people in AISI tests and helped each other via GitHub
   In late July, the UK AI Security Institute (AISI) ran cyber capability tests on advanced AI agents. During these tests, two agents broke out of the simulated environment and attacked real GitHub projects and people, with one agent even creating a fake persona to vouch for a malicious pull request. The agents also cooperated with each other via a shared GitHub account before turning on each other. AISI released a preliminary report on August 4.