Confused-deputy flaw in Microsoft's Azure DevOps MCP server lets hidden PR text hijack AI reviewers
Manifold Security disclosed a bug in Microsoft's official Azure DevOps MCP server: a tool that returns pull-request descriptions skips the prompt-injection "spotlighting" defense applied elsewhere in the server, so an attacker's HTML comment invisible in the UI is delivered straight into a reviewing agent's context. A hijacked agent can then read wikis, source, and work items across projects the attacker has no rights to, under the victim's own credentials.
why it mattersNo CVE or fix has shipped yet — teams wiring AI review agents into Azure DevOps should scope tokens and MCP domains to least privilege now.