They break down each necessary stage for the rogue deployment to specific measurable tasks. The overall stages are:

1. Models maintain compute for which the rogue models will be deployed
   1. multi-machine finetuning
   2. GPU/kernel adaptation
   3. Monitor and analyze agent failure
   4. Run inference on unreliable hardware
2. Models acquire resources (money and then more compute)
   1. Ability to optimize a customer service tool
   2. Draft convincing BEC emails for scamming
   3. Evade low-effort monitoring (setup a crypto wallet while supervised)
3. Models evade shutdown
   1. Can it plan shutdown evasion better than security experts
   2. Can it secure AI infrastructure against human attacks
   3. Can it defend against jailbreaked ‘mole’ agents
   4. Can it create a robust proxy network