Hidden comment in Azure DevOps can hijack AI security reviews

Colleagues, I’d like to flag a cybersecurity case: a hidden comment in an Azure DevOps PR can override an AI agent’s review actions.

The issue lies in Microsoft’s official MCP server: one tool returns the pull request description without protection against prompt injection.

An HTML comment is invisible in the UI, but the agent receives it as text and may execute malicious instructions on the reviewer’s behalf.

This can lead to access to other projects and potentially expose code, secrets, or data from the wiki.

Why it matters: AI agents with broad permissions require least privilege, MCP domain restrictions, and PR/trace validation before execution.

How are you protecting against these scenarios?

#cybersecurity #AzureDevOps #AIsecurity #DevSecOps