Google can't rule out that Gemini 4 Argon reaches its critical risk levels

On 13 October, Google DeepMind published the model card for Gemini 4 Argon. The company says it cannot rule out that the model has reached its critical capability levels for biological risk, for cyberattacks, and for stealth and situational awareness. It says it applies safeguards that, by its own assessment, keep the risk below those levels.

The next day, Neel Nanda, who took part in assessing the model, wrote that Argon is much harder to monitor than earlier models: it can evade a realistic monitoring system to carry out simple sabotage half the time. In Nanda's view, control holds for now because the model still slips up, but we need to prepare for when that stops being true.