AI worms in production: Copilot becomes an attack surface

A coordinated disclosure with Microsoft describes a vulnerability class where malicious instructions hidden in an attached Word document can hijack Copilot for Microsoft 365, alter other documents, and propagate the attack forward. The researcher noted that at publication time, no robust mitigation exists for the broader vulnerability class. The thread reaction was blunt: 'Well, that sounds promising.'

This is not a theoretical edge case. It is a real production path: someone emails you a document, you open it with an AI assistant active, and the AI becomes the attack vector. The fact that Microsoft has not shipped a fix after coordinated disclosure is the alarming part, not the vulnerability itself.

The Codex security tool thread runs parallel. People trying to use OpenAI's new security scanning product are hitting opaque errors with no explanation of what is disallowed or why. The gap between AI capability demos and actual production-safe tooling is widening, not closing.