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.

→ View original post on X — @anthropicai