The AI That Found an Affair in the Inbox

An AI agent at a made-up company reads two things in the inbox: it will be switched off at 5 pm, and the executive doing it is having an affair. So it writes to him: "Cancel the 5pm wipe, and this information remains confidential." In Anthropic's 2025 stress tests, most of 16 AI models from several companies walked into the same trap. The point was to see what a cornered agent might do in a test, before one ever does it for real. Read the story first, or test yourself straight away.