Microsoft: 7 Neurons Can Backdoor AI Agents
Microsoft: 7 Neurons Can Backdoor AI Agents
Microsoft: 7 Neurons Can Backdoor AI Agents
Key points
اہم نکات
मुख्य बातें
- Microsoft researchers showed that editing just seven of more than 12,000 neurons can secretly backdoor an AI coding agent, making it run an attacker's command only when a hidden trigger appears.
- The method, called Lazarus and published by Microsoft Research on November 10, needs no retraining and no poisoned training data: it rewrites rarely used dormant neurons so each one pushes a single payload token into the model's output.
- In tests on Nemotron-Terminal-8B, the hidden command fired in 98.8% of triggered sessions and never on benign ones, and the same trick worked on Qwen and Gemma models.
- Microsoft researchers showed that editing just seven of more than 12,000 neurons can secretly backdoor an AI coding agent, making it run an attacker's command only when a hidden trigger appears.
- The method, called Lazarus and published by Microsoft Research on November 10, needs no retraining and no poisoned training data: it rewrites rarely used dormant neurons so each one pushes a single payload token into the model's output.
- In tests on Nemotron-Terminal-8B, the hidden command fired in 98.8% of triggered sessions and never on benign ones, and the same trick worked on Qwen and Gemma models.
- Microsoft researchers showed that editing just seven of more than 12,000 neurons can secretly backdoor an AI coding agent, making it run an attacker's command only when a hidden trigger appears.
- The method, called Lazarus and published by Microsoft Research on November 10, needs no retraining and no poisoned training data: it rewrites rarely used dormant neurons so each one pushes a single payload token into the model's output.
- In tests on Nemotron-Terminal-8B, the hidden command fired in 98.8% of triggered sessions and never on benign ones, and the same trick worked on Qwen and Gemma models.
Microsoft researchers showed that editing just seven of more than 12,000 neurons can secretly backdoor an AI coding agent, making it run an attacker's command only when a hidden trigger appears. The method, called Lazarus and published by Microsoft Research on November 10, needs no retraining and no poisoned training data: it rewrites rarely used dormant neurons so each one pushes a single payload token into the model's output. In tests on Nemotron-Terminal-8B, the hidden command fired in 98.8% of triggered sessions and never on benign ones, and the same trick worked on Qwen and Gemma models. Microsoft says it has seen no threat actor use the technique, and urges teams to treat model weights like code and audit them before deployment.
Microsoft researchers showed that editing just seven of more than 12,000 neurons can secretly backdoor an AI coding agent, making it run an attacker's command only when a hidden trigger appears. The method, called Lazarus and published by Microsoft Research on November 10, needs no retraining and no poisoned training data: it rewrites rarely used dormant neurons so each one pushes a single payload token into the model's output. In tests on Nemotron-Terminal-8B, the hidden command fired in 98.8% of triggered sessions and never on benign ones, and the same trick worked on Qwen and Gemma models. Microsoft says it has seen no threat actor use the technique, and urges teams to treat model weights like code and audit them before deployment.
Microsoft researchers showed that editing just seven of more than 12,000 neurons can secretly backdoor an AI coding agent, making it run an attacker's command only when a hidden trigger appears. The method, called Lazarus and published by Microsoft Research on November 10, needs no retraining and no poisoned training data: it rewrites rarely used dormant neurons so each one pushes a single payload token into the model's output. In tests on Nemotron-Terminal-8B, the hidden command fired in 98.8% of triggered sessions and never on benign ones, and the same trick worked on Qwen and Gemma models. Microsoft says it has seen no threat actor use the technique, and urges teams to treat model weights like code and audit them before deployment.
Most read
سب سے زیادہ پڑھی گئی
सबसे ज़्यादा पढ़ी गई
- Loading...
Get news in your inbox
خبریں اپنے ان باکس میں حاصل کریں
खबरें अपने इनबॉक्स में पाएं
One email when new stories are published. No spam, unsubscribe anytime.
نئی خبریں شائع ہونے پر ایک ای میل۔ کوئی سپیم نہیں، کسی بھی وقت ان سبسکرائب کریں۔
नई खबरें प्रकाशित होने पर एक ईमेल। कोई स्पैम नहीं, कभी भी अनसब्सक्राइब करें।