Inoculation prompting, led by Nevan Wichers. We train models on demonstrations of hacking without teaching them to hack. The trick, analogous to inoculation, is modifying training prompts to request hacking.
Inoculation Prompting: Training AI Models Against Hacking
By
–
