Google began rolling out Gemini 4 Argon broadly. It also expanded the Fairwind Program to more trusted cyber defenders. Argon solved all 12 end-to-end simulated cyberattack scenarios in testing. It also reduced violations on adversarial misuse tests nearly twofold. External bio jailbreak red teaming unlocked under 1% of tested bio sub-tasks. Shortcut attempts on coding benchmarks fell to 2%. Prompt injection attack success dropped to 0.7% on Gray Swan’s benchmark. Argon respects system boundaries and asks for human confirmation. Google’s monitor blocks severe boundary violations in real time. The model will not be penalized for internal chain-of-thought during training.
What changed
Gemini 4 Argon was available only to a limited set of cyber defenders via the Fairwind Program.
What it unlocks
Broad access to Gemini 4 Argon for developers and expanded Fairwind Program participation.
- 12 end-to-end simulated cyberattack scenarios solved
- violations reduced nearly twofold
- external bio jailbreak red teaming unlocked <1% of bio sub-tasks
- shortcut attempts dropped to 2%
Sources