Google’s October model card says it cannot rule out that Gemini 4 Argon has reached the company’s capability thresholds for biological information risk or cybersecurity. Google says safeguards reduce its assessed deployment risk below those levels; the card does not report that the model caused a biological or cyber incident.
A Critical Capability Level (CCL) is a capability threshold at which a frontier model could pose a heightened risk of severe harm if it had no mitigations. Google’s phrase “cannot rule out” means the company cannot exclude that Argon has reached the relevant capability level. It is not a declaration that the model definitely crossed it.
For biological information, the card says expert red-team evaluations showed measurable capability uplift over web baselines and previously tested models in two scenarios. It describes Argon’s responses as comprehensive, accurate and actionable, then says its safety case assesses the risk below the biological CCL after safeguards.
For cybersecurity, Google cites substantial uplift in its evaluations and quantitative modeling, and says severe harm could be possible without adequate safeguards. It says a safety case assesses the remaining risk below the cybersecurity CCL. These are the company’s assessments of its own controls.
The card also says Google cannot rule out a Tracked Capability Level (TCL) for stealth and situational awareness after Argon passed 14 of 15 challenges. Google says safeguards reduce that risk below the TCL. The framework describes TCLs as earlier signals for tracking potential risks that may be less severe than CCL-level risks. The same card says Argon has not reached the CCL for automating machine-learning research and development.
That distinction is central to the disclosure: Google says the model’s unmitigated capabilities may sit at several thresholds, while its post-mitigation risk assessments fall below them. The public model card presents those mitigations as Google’s own conclusions; it does not establish how they perform across every real-world use.
Illustration by jaydeep_ via Wikimedia Commons, dedicated under CC0 1.0. Generic cybersecurity artwork, not a depiction of Gemini 4 testing.
Sources
- Google DeepMind, Gemini 4 Argon Model Card, published Oct. 13, 2026.
- Google DeepMind, Frontier Safety Framework update, updated Apr. 17, 2026; defines CCLs and explains the earlier-warning role of TCLs.