An email should not decide what an assistant reveals

What was reported

Microsoft describes EchoLeak, CVE-2025-32711, as a now-fixed prompt-injection vulnerability in Microsoft 365 Copilot. Under certain conditions, a crafted email could redirect the assistant into exposing limited internal information that the user already had permission to access. This briefing concerns a disclosed vulnerability and research demonstration, rather than a confirmed customer breach.

Read the report from Microsoft

Microsoft’s vulnerability advisory: CVE-2025-32711

Our perspective

Our reading: permission to retrieve a document does not make that document a trusted source of instructions. Task-scoped context and instruction boundaries address different questions. This case informs Vivendur’s thinking about access without implying that a memory-permission system alone solves prompt injection.