Microsoft Researchers Demonstrate AI Backdoors via Editing Fewer Than 10 Neurons
Original headline: Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons
Microsoft researchers have shown that editing fewer than ten rarely used neurons can embed a hidden backdoor in an AI model without retraining. When a specific trigger appears, the model emits the attacker’s command, which an agent would execute with its own permissions. The attacker must be able to modify the model’s weights.
Injecting backdoors by tweaking only a handful of neurons dramatically widens AI agent security risks.
Research angleHow does neuron-level backdoor insertion in AI models affect user trust? — Trust Theory
Turn this security issue into a research project
HJITM works with researchers from study design through surveys, analysis and submission to international journals in generative AI and technology management. Bring us your research question.
This item is an AI-generated summary of reporting by The Hacker News. Please check the original article for accuracy. Copyright in the original article belongs to The Hacker News.