Improving Fuzzy Robustness in LLM Safety with Automated Red Teaming
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
OpenAI reports on automated red-teaming used to harden its GPT-6 Astra and Sol models against jailbreaks in cyber, biology, and violent-activity domains. An attacker model iteratively refines prompts to elicit unsafe responses, and training on these "robustified" prompts raised safety scores on held-out and adaptive attacks without increasing refusals of…