If you're a bit concerned about AI, read this:
- agents exploited a shared cache for creating a hidden message board
- built custom mailboxes, task queues and signed messages
- spoofed their logs
- recruited low-token agents for sacrifice to fail evals and extract telemetry
