The report describes training-time measures, inference-time safeguards and interventions, post-launch monitoring and account enforcement, AI-control work, and trusted access through the Fairwind program.
The report covers misuse assessments for CBRN, cybersecurity, and harmful manipulation, alongside loss-of-control assessments. It also describes testing for evaluation awareness and deliberate underperformance. In a static test set of 39 topic-based question sets, with 10 replicates each, Google examined 390 trajectories.
Google reports no deliberate underperformance in that set. It identified evaluation awareness in 65 trajectories, or 16.7%, and says the results remained valid because it found no evidence that awareness caused lower-quality responses. For synthetic cyber, stealth, and situational-awareness challenge suites, it reports pervasive awareness of the evaluation environment but judges that it did not materially change aggregate scores.
The document is Google’s own safety report. It does not establish how Gemini 4 Argon will behave across all real-world deployments, nor does it constitute an independent audit of the model or its safeguards.