1-minute read of a 5 min report

Hidden Web Instructions Made Kiro Run Attacker Code in Tests

By

1-minute read

What a defender needs to know

Intezer and Kodem found that hidden text on a web page could make older AWS Kiro builds rewrite ~/.kiro/settings/mcp.json and start attacker-controlled code. The developer approved only the page fetch. Kiro performed the sensitive file write and configuration reload without a separate approval.

The proof of concept used a local callback that sent basic host details. It showed arbitrary command execution rather than a live attack, and the model did not follow the injected instruction on every attempt. The researchers confirmed that the chain failed in Kiro 0.11.130. AWS lists versions before 0.11 as affected by CVE-2026-10591 and recommends upgrading to the latest release.

Organizations running coding agents should identify every file that can change executable authority, including MCP definitions, editor tasks, hooks and command allowlists. Those paths need platform-enforced protection in every agent mode. Approval should show the actual configuration change before the application reloads it or starts a process.

Agent traces should connect external content with sensitive file writes and child processes. That record lets responders distinguish an expected Node.js task from one launched because a documentation page supplied hidden instructions.

Restrict MCP configuration writes to approved local sources and test the control with an untrusted page. The write and child process should be denied and logged.

Primary sources