Google DeepMind rolls out Gemini 4 Argon with cyberdefense and safety safeguards

Google DeepMind is beginning to roll out Gemini 4 Argon broadly after a limited access period, and expanding its Fairwind Program to more trusted cyber defenders.

The model solved all 12 end-to-end simulated cyberattack scenarios, up from five in the previous generation.

Google describes three real-time protection layers against misuse, a misalignment monitor, and prompt injection resistance, reporting that attack success on Gray Swan's IPI benchmark fell to 0.7% across 15 adaptive attempts.