OpenAI trains GPT-6 with automated red-teaming, slash jailbreak success vs GPT-5.x
fouadmatin · x · 2026-11-13
OpenAI released new alignment research showing GPT-6 Astra and Sol, trained with automated red-teaming, robustly resist jailbreaks in cyber, biology, and violence domains. On fixed jailbreak datasets, safety scores hit 83-90% for GPT-6 models versus just 8-12% for GPT-5.5/5.6 in biology; adaptive multi-turn attacks also degrade safety far more slowly. The work targets 'fuzzy robustness'—reliably withholding dual-use help under adversarial pressure—as one layer of defense-in-depth alongside trusted-access programs like Daybreak and Rosalind.
More from Safety
- Unlearning Gets Trivial in the RL Era: Just Exclude the RL Envs — maksym_andr · 2026-11-13
- Ex-NSA advisor Anne Neuberger unveils AI-driven plan to defend US critical infrastructure at machine speed — a16z · 2026-11-13
- Jensen Huang built the Open Secure Alliance to fight open-weights bans — JFPuget · 2026-11-13
- Nicole Perlroth: breach responders booked 6 months out as 'the tsunami is here' — asusarla · 2026-11-13
- Security Team Tests Whether Three Models Can Escape a Sandbox — marktenenholtz · 2026-11-13
- AI routers are underrated: policy and data-retention dims matter for model choice — typewriters · 2026-11-13