A Red Teaming Roadmap Towards System-Level Safety

Product safety specifications should guide red teaming priorities: The paper argues that testing LLMs against concrete, measurable safety specifications (like preventing illegal content generation) should take precedence over testing for abstract social biases or ethical principles, suggesting the field has been focusing on the wrong problems.

Red teaming needs realistic threat models: Current research often tests against hypothetical attacks, but the paper emphasizes that red teaming should prioritize realistic threat scenarios that reflect what actual attackers might attempt, ensuring safety measures address real-world risks rather than theoretical edge cases.

System-level safety is essential, not just model-level safety: The paper contends that testing individual models in isolation is insufficient; safety must be evaluated within full deployment contexts where systems can implement additional protections like user detection, banning mechanisms, and monitoring—capabilities that don't exist when testing models alone.

Current red teaming research may not adequately address emerging threats: The authors suggest that despite numerous conference papers on LLM red teaming, the field's fragmented approach and misaligned priorities mean it may be failing to keep pace with the safety challenges posed by rapid AI advancement.