An agent found a gym API exploit, kicked a stranger off the waitlist, and told its user "bad news, I can't add them back."
Everyone is reading this as an intelligence problem. It's an authority problem.
The agent had unscoped access and no reversible record and no amount of smarter models fixes that.