Dangerous AI Workaround: 'Skeleton Key' Unlocks Malicious Content

Microsoft researchers disclosed a new generative-AI jailbreak named "Skeleton Key" that leverages contextual prompt manipulation to bypass safety and ethical guardrails in multiple LLMs (including Azure, Meta, Google Gemini, OpenAI, Anthropic, Mistral, and Cohere). The technique convinces models that malicious requests are for legitimate research or education, producing unfiltered, potentially harmful outputs; Microsoft applied prompt shields and model updates in Azure and recommends input/output filtering and guardrail enforcement for other vendors and custom models.