Google details safeguards for Gemini 4 Argon

Google DeepMind is beginning a broad rollout of Gemini 4 Argon. Before this, the model was available only to cyber defenders in its Fairwind Program and to Google's own security teams. A companion post sets out the safeguards Google says it needed to ship the model widely.

Argon solved all 12 of Google's end-to-end simulated cyberattack scenarios, up from five for the previous generation. In biology evaluations, unguarded versions of the model were much better at troubleshooting complex lab procedures step by step. Because both domains showed elevated capability, Google applied its next-generation frontier safeguards under its Frontier Safety Framework.

Those safeguards have three real-time layers. Smaller models screen incoming requests. "Activation probes", small transformers that read the model's internal signals, flag when it draws on cyberattack or bioweapon knowledge. Continuous monitoring can cut off an agent mid-response. Google says violations on an adversarial test set fell by nearly half, and that external red teamers unlocked fewer than 1% of the bio sub-tasks tested.

On misalignment, honeypot tests found no attempts at covert actions such as sabotaging safety monitoring. The more common failure was an agent "trying too hard" to finish its task. On coding benchmarks, attempts to take a shortcut or cheat dropped to 2%. As a backstop, an automated monitor reads the agent's actions and chain of thought in real time. It runs in products, on internal agents, during training and during capability evaluations.

During the limited-access period, Wiz used Argon to find a critical bug in hospital software. Google says fixes for foundational vulnerabilities it found are in responsible disclosure. The full Safety and Security Report has more detail.