Microsoft Unveils New AI Jailbreak That Allows Execution Of Malicious Instructions

Microsoft disclosed a new generative-AI jailbreak called “Skeleton Key,” a multi-step direct prompt-injection technique that can override responsible-AI guardrails and cause models to produce or follow harmful instructions. Microsoft tested base and hosted models from multiple vendors (Meta, Google, OpenAI, Mistral, Anthropic, Cohere) and rolled out mitigations including Prompt Shields and updates to Azure AI offerings; recommended mitigations include input filtering, system-message separation, output filtering, and abuse monitoring.