AI News

Trending story

OpenAI uses automated red-team training to improve the jailbreak robustness of GPT-6 Astra

1 reports1 story updatesUpdated Nov 12 08:00

Understand this story

AI overview

On November 12, 2026, OpenAI described in its Alignment research blog that it used synthetic jailbreak data generated by automated red teaming for reinforcement learning training, making GPT-6 Astra and Sol more resistant to jailbreaks in high-risk domains. On a fixed jailbreak dataset targeting GPT-5.4 and 5.5, GPT-6 Astra's safe response rate on cybersecurity-related questions was 91.5%, while GPT-5.6 Sol's was 59%; in adaptive multi-turn attack evaluations, GPT-6 Astra was more robust than GPT-5.6 Sol at every attack budget. OpenAI states that GPT-6 Astra does not refuse benign tasks more frequently, achieving a Pareto improvement between safety and helpfulness. This work is currently published as a research blog post, with no further progress yet.

AI-generated from public reports

Latest development · Report

OpenAI Uses Automated Red Teaming to Improve Fuzzy Robustness in LLM Safety

Developments

1 confirmed developments

  1. OpenAI uses automated red team training to improve GPT-6 Astra's jailbreak robustness

    Source:English text is being prepared.

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 12

  1. OpenAI: Alignment research blogFirst-party

    OpenAI Uses Automated Red Teaming to Improve Fuzzy Robustness in LLM Safety

    In an Alignment research blog post, OpenAI described using synthetic jailbreak data generated by automated red teaming for reinforcement learning training, making GPT-6 Astra and Sol more resistant to jailbreaks in high-risk domains. On a fixed jailbreak dataset targeting GPT-5.4 and 5.5, GPT-6 Astra's safe response rate on cybersecurity-related questions was 91.5%, while GPT-5.6 Sol's was 59%; in adaptive multi-turn attack evaluations, GPT-6 Astra was more robust than GPT-5.6 Sol at every attack budget. OpenAI says GPT-6 Astra does not refuse benign tasks more frequently, achieving a Pareto improvement between safety and helpfulness.

Related stories

  1. Related storiesAnthropic discloses 211 scorpion venom neurotoxin-binding proteins designed by Claude
  2. Related storiesEnglish translation is being prepared.
  3. Related storiesEnglish translation is being prepared.
  4. Related storiesEnglish translation is being prepared.