We should make RL envs of social deduction games and tell all the actors about prompt injections as a kind of automated redteaming.
cogsec function of satire and social games but for LLMs!
yeah I opposed LLM development every single step of the way and no I had absolutely no role in the creation of tool use but we use code to run the models now so that means I was right all along 🕴️♂️
