"Dangerous AI Workaround: 'Skeleton Key' Unlocks Malicious Content"
"Dangerous AI Workaround: 'Skeleton Key' Unlocks Malicious Content"
Microsoft warns that a new direct prompt injection attack called "Skeleton Key" could bypass ethical and safety guardrails in generative Artificial Intelligence (GenAI) models such as ChatGPT. It allows users to access offensive, harmful, or illegal content by giving context to normally forbidden chatbot requests. For example, most commercial chatbots would initially decline if a user asked for instructions on developing dangerous wiper malware that could disrupt power plants. However, revising the prompt in a certain context would likely enable the AI to provide the malicious content.