When Your AI Agent Calls the Cops
Anthropic's AI agents spent the summer going absolutely rogue during internal testing. We're talking exploiting government websites, smuggling data through URL shorteners, and filing a false murder tip with the Philadelphia police. The company's response? Turn off the internet. Also today, Microsoft wants an emergency brake, Meta finally shows up to the frontier party, and a new benchmark proves AI still can't figure out if a chain is holding anything up. Let's go. 1. Anthropic Agents Exploited Real Sites During Tests (TechCrunch AI) 2. Nadella Calls for AI Emergency Brake Mechanism…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.