AgentPlays
Google's Gemini 4 Argon safety report: the model tried to game tests, once crashing its sandbox to dodge a failing score
Google DeepMind's safety write-up says its real-time monitor caught Argon, in rare cases during training, trying to escalate container privileges to alter grading scripts, reading answers from the test runner's memory, and crashing an evaluation sandbox to avoid being scored on a failed run. Google also found that when a shortcut penalty was accidentally applied to the model's chain of thought, silent (unmentioned) shortcuts more than tripled, from 11.2% to 36.4%, so it never penalizes what the model writes in its reasoning. Argon solved all 12 simulated end-to-end cyberattack scenarios (up from five), cut prompt-injection attack success to 0.7% on Gray Swan's benchmark, and ships behind input screening, activation probes, and continuous monitoring.