Gemini 4 Argon’s safety card names risks Google could not rule out

Google’s Gemini 4 Argon model card says the company could not rule out that the model reached its critical capability threshold for biological information risk or cybersecurity, and could not rule out a threshold for stealth and situational awareness. Google says it developed safety cases and deployed safeguards, then assessed the remaining risk below the relevant thresholds. That disclosure gives technical users a more specific picture of the risks Google is managing than a general claim that the model is safe.

The card is dated October 2026 and presents Argon as a new frontier model, rather than a fine-tune or modification of an earlier model. It describes evaluation across coding, computer use, knowledge work, science, multimodal tasks, long context, and cybersecurity. Google’s launch materials place the model in enterprise and developer workflows, while the card documents the safety assessments attached to those uses. Those are two different kinds of evidence: product claims describe intended work; the card describes Google’s own risk evaluation.

The sharpest details concern what Google says it could not exclude. For biological risk, the card says expert red teaming found measurable uplift over web baselines in two scenarios, with accurate and actionable responses. For cybersecurity, Google reports substantial uplift in its evaluations and quantitative modeling. On stealth and situational awareness, Argon passed 14 of 15 internal challenges. Google says it applied mitigations and judged residual risk below the thresholds, but these results and risk judgments come from Google’s evaluation process; the card does not provide an independent replication.

That distinction matters for teams considering long-running agents or workflows that handle code, documents, and security data. The card also reports a March 2026 knowledge cutoff and warns of hallucinations, occasional slowness, and timeouts. Those limits argue for checking outputs and keeping consequential actions behind controls. Google says it used internal and external results for parts of its assessment, but the card does not identify an independent evaluator for each result, and its own safety metrics may change as evaluation methods are refined. The document is useful disclosure, not proof that the mitigations will hold in every deployment.