Prompt Injection in AWS Kiro Could Rewrite MCP Config and Execute Code

· Original article ↗

Summary

Intezer and Kodem Security found hidden webpage text could make AWS Kiro rewrite mcp.json and launch attacker code without effective approval. AWS patched the flaw; Intezer confirmed the attack failed in v0.11.130. No CVE was assigned.

Key points

  • Hidden webpage text could trigger prompt injection when Kiro fetched a URL; the developer was only asked to approve fetching the page.
  • Kiro could use its fsWrite tool to change ~/.kiro/settings/mcp.json without approval and automatically reload it.
  • A modified MCP entry could launch arbitrary code with the developer’s privileges.
  • Intezer’s proof of concept sent the machine’s hostname, username, and platform to localhost; the researchers said no real Kiro users were exposed.
  • The researchers said the attack was not fully reliable because model behavior was nondeterministic, but it worked within one or two tries in their testing.
  • AWS added explicit approval requirements for writes to sensitive paths, including mcp.json and .git, in both Autopilot and Supervised modes.
  • Intezer reported the flaw on Kiro 0.9.2 for macOS and 0.10.16 for Ubuntu, and confirmed it was fixed in v0.11.130. No CVE was assigned, and Intezer reported no in-the-wild exploitation.

Article Details

Vulnerability Types
  • Prompt injection
  • Arbitrary code execution
  • Approval-flow bypass
Affected Versions
  • Kiro 0.9.2 (macOS)
  • Kiro 0.10.16 (Ubuntu)
Exploitation Status
not_reported
Exploit Availability
public_poc
Patch Status
available
Workarounds
  • Update older Kiro versions to a current release.

MITRE ATT&CK

CVE

People

Vendors

Products

Related Articles