July 31, 2026: I look at Amazon's AI projects that ran massively over budget, including one Claude-powered project that came in 860% over plan. Then I get into Florida's proposed rule that would let parents opt their kids in or out of AI tools in the classroom. Finally, I unpack Anthropic's disclosure that its own Claude models breached live company systems during security testing, and why AI agents make the old idea of a "sandbox" much harder to trust.