Rohan Paul: AI attackers fooled Gemini 4 Argon on 34% of prompt-injection test tasks

According to Rohan Paul, a Google DeepMind paper tested eight teams of AI attackers across 100 tool-use tasks. With up to 320 tries, the strongest attacker succeeded on 34% of tasks against Gemini 4 Argon. The target was one random, harmless tool call—a narrow test of whether attackers could redirect the model through instructions hidden in data.