Google rolls out Gemini 4 Argon broadly after a limited-access window for cyber defenders

Google ties the defensive and offensive results together itself. The same coding and reasoning improvements that let Argon find and patch vulnerabilities also drove its cybersecurity test scores ^[4]. Against the previous generation, that is seven more simulated attack scenarios solved ^[13]. The scenarios are simulated and Google's own, so the score predicts a real network only if that network looks like the scenarios. A suite at 12 of 12 also has no room left to show the next gain ^[3].

Activation probes read the model's internal signals. They are trained to flag the moments when the model draws on what it knows about cyberattacks or biological threats ^[7]. Reading activations requires access to them, so we'd expect the probes to run only where Google controls the serving stack. The third layer combines those signals across a multi-step response and can cut off an agent mid-stream ^[8].

A long scanning job that handles exploit code may look, to a classifier trained on cyberattack knowledge, like the thing it is built to stop. Google says the layers are aimed at attackers who disguise a request as a harmless story or break it into dozens of small questions over several weeks ^[9]. We think a multi-week audit can resemble that pattern at the request level.

The bio jailbreak figure depends on its denominator. External red teaming could unlock less than 1% of the bio sub-tasks Google tests for ^[10]. That is a percentage of Google's own sub-task list, so it tells a buyer how the gate performs on Google's threat model.

That previous models had missed the hospital bug is Google's statement ^[14]. The post does not name the hospital product, its vendor, or the foundational vulnerabilities, which affect systems globally and are going through responsible disclosure ^[16].