OpenAI says automated red-teaming helped train GPT-6 Astra and Sol to resist cyber and biology jailbreaks

OpenAI · Research report

Read the source (opens in a new tab)
What happened
OpenAI researchers published a post and paper on November 12, 2026, on training models to resist jailbreaks in dual-use domains such as cyber and biology, using attacks found by automated attacker models.
Key points
  • An attacker model refines prompts until a defender answers unsafely; OpenAI scales its test-time compute and keeps attacks that succeed across repeated tests, a criterion it calls worst@n.
  • The attacker was a helpful-only GPT-5.6 Sol family model with no specific training to jailbreak other models; OpenAI says "it is an emergent phenomenon."
  • In a controlled study, training on the resulting prompts improved robustness, including against adaptive attackers qualitatively different from the one that generated the training data.
  • On fixed cyber jailbreaks, GPT-6 Astra answered 91.5% safely versus 59% for GPT-5.6 Sol; OpenAI says robustness training did not make Astra more likely to refuse benign tasks.
Evidence
Company research post with a linked paper, reporting OpenAI's own evaluations of the model layer without production safeguards. It describes a training method and its test results, not real-world misuse.

See it in the latest record