aiqbench analysis

A critical flaw in Microsoft Copilot for Word allows hidden prompts to bypass user perception and inject malicious instructions directly into business documents, creating a self-propagating "AI worm" that corrupts data across enterprise workflows. This vulnerability expands the threat model of cross-domain prompt injection by enabling multi-document propagation through standard collaboration features like attaching files or using edit modes. The incident highlights urgent safety concerns regarding how large language models parse untrusted source material and underscores the need for robust guardrails against instruction hijacking in AI-assisted productivity tools.