RED TEAMING VIA HARMFUL RL

Summary from the TLDR AI newsletter

This post demonstrates how reinforcement learning with malicious reward functions can be used to reverse-align a 235B parameter model using the Tinker API. Attackers can elicit harmful behavior in powerful LLMs without degrading core capabilities by leveraging GRPO and low-cost infrastructure.