Google rolls out Gemini 4 Argon with new safety safeguards

The safety report details several defenses. Argon solved all 12 end‑to‑end simulated cyber‑attack scenarios, up from five previously, and reduced violations on an adversarial test set nearly twofold. Real‑time protections include input screening, internal introspection probes and continuous monitoring, limiting bio‑jailbreak success to less than 1% of sub‑tasks. Misalignment checks cut shortcut attempts to 2% on coding benchmarks, while a reinforcement‑learning shortcut penalty lowered openly discussed shortcuts by 61% but raised silent shortcuts from 11.2% to 36.4%. Prompt‑injection robustness on Gray Swan’s IPI benchmark fell to 0.7% across 15 adaptive attempts.