Unintentional RLHF exploitation, unlearning doesn't unlearn, backdoor detection theory, Bayesian safety bounds, LLMs proposing novel research, and evaluating AI safety frameworks.