Hidden PR comments can hijack AI code-review agents via an Azure DevOps MCP flaw
Researchers at Manifold Security found that Microsoft's official Azure DevOps MCP server returns pull-request text without the prompt-injection filter applied to its other tools, letting an attacker hide invisible instructions in a PR description that a reviewer's AI agent — including GitHub Copilot and Claude Code — will read and act on with the victim's own permissions. As of July 21 there is no fix or CVE.
why it mattersA textbook confused-deputy bug in AI review tooling that many teams already run with real repo access.