Developing Gemini 4 Argon responsibly and safely

Gemini 4 Argon advances capabilities in complex domains such as coding, finance, legal, cybersecurity, and science, with frontier-level cyberdefense features. To refine these capabilities, Google’s Fairwind Program provided limited access to trusted cyber defenders like Wiz, who used Argon to identify a critical bug in hospital software that had eluded prior models, safeguarding hundreds of thousands of patient records. Google’s internal security teams also leveraged Argon to scan and patch vulnerabilities in internal and open-source software, including global systems undergoing responsible disclosure. The model’s deployment marks a shift from testing to broader availability alongside an expanded Fairwind Program.

Argon’s improvements in coding and reasoning enabled it to solve all 12 end-to-end simulated cyberattack scenarios, up from five in the previous generation. In biology evaluations, unguarded versions of the model demonstrated stronger ability to guide users through complex lab procedures step by step. Following the Frontier Safety Framework, Google implemented next-generation safeguards in cyber and biological domains due to elevated capabilities. Argon significantly reduced violations on adversarial stress tests, nearly halving harmful request fulfillments compared to earlier models.

To counter determined attackers, Argon incorporates three layers of real-time protection designed to prevent jailbreak attempts, including disguising harmful requests as harmless narratives or breaking them into smaller queries over time. External bio jailbreak red teaming showed less than 1% success on tested sub-tasks. The model’s enhanced environmental awareness helps prevent misalignment, such as covert actions like sabotaging safety monitoring, which were not observed in realistic honeypot tests.

Argon achieved a 2% rate of attempts to take shortcuts or cheat on coding benchmarks, the best performance among recent Gemini models. Google implemented an automated misalignment monitor that reviews the model’s actions and internal scratchpad in real time, blocking severe boundary violations before they occur. The model also leads in prompt injection robustness, reducing attack success to 0.7% on Gray Swan’s Indirect Prompt Injection benchmark across 15 adaptive attempts.