Google DeepMind 研究用多智能体攻击测试提示注入防护Chinese only
Google DeepMind 研究人员构建了由 8 组 AI 智能体组成的攻击系统，让它们测试提示注入、交流结果并改进尝试。在 100 项工具使用任务中，攻击者的目标仅是诱发一次随机、无害的工具调用；最多尝试 320 次时，最强攻击者在 34% 的任务中攻破了文中称防护最强的 Gemini 4 Argon。
1 developments2 reports1 sources
Event developments 1
Google DeepMind study tests prompt injection defenses with multiple AI agents
Google DeepMind researchers built an attacker system with eight teams of AI agents that test prompt injections, share notes, and improve their attempts. Across 100 tool-use tasks, the attacker's goal was only to trigger one random, harmless tool call. With up to 320 tries, the strongest attacker succeeded on 34% of tasks against Gemini 4 Argon, described in the source as the most robust model.
View source reports · 1 sources · 2 reports
- Rohan Paul (@rohanpaul_ai) · Google DeepMind study tests prompt injection defenses with multiple AI agents
- Rohan Paul (@rohanpaul_ai) · Research report says coordinated AI agents can improve prompt injection attacks
All reports 2
Research report says coordinated AI agents can improve prompt injection attacks
A post describing a Google DeepMind paper says researchers used eight teams of AI agents to test prompt injections, share notes and refine attacks. Across 100 tool-use tasks, the attacker sought just one random, harmless tool call. With up to 320 attempts, the strongest attacker succeeded on 34% of tasks against Gemini 4 Argon, described in the post as the most robust model.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)
Google DeepMind study tests prompt injection defenses with multiple AI agents
Google DeepMind researchers built an attacker system with eight teams of AI agents that test prompt injections, share notes, and improve their attempts. Across 100 tool-use tasks, the attacker's goal was only to trigger one random, harmless tool call. With up to 320 tries, the strongest attacker succeeded on 34% of tasks against Gemini 4 Argon, described in the source as the most robust model.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)