OpenAI's GPT-6 Astra hits Critical cyber capability tier with near-99.9% prompt injection robustness

morqon · x · 2026-09-05

OpenAI released the system card for GPT-6 Astra, its most capable broadly deployed model to date:

First model to reach the Critical tier of the Preparedness Framework for cybersecurity — with the right tools it can find previously unknown vulnerabilities and develop new exploits across well-protected systems without step-by-step human guidance.
Hardened internals: stricter isolation, checkpoint encryption, universal monitoring of full trajectories including chains of thought, and blocking alignment evaluations before internal use.
Jailbreak robustness: new robustness safety training makes Astra significantly more resistant than GPT-5.6 Sol across longer trajectories, backed by internal/external red-teaming, regression testing, and conservative refusal boundaries for high-risk users.
Commentary notes Astra approaches 99.9% prompt injection robustness, seen as a milestone for deploying agents to a billion users.

More from Models