Varonis Details Dialogflow CX Flaw That Could Let Attackers Hijack AI Conversations

· Original article ↗

Summary

Varonis reported that a Dialogflow CX flaw could let attackers alter Code Blocks to access conversations, manipulate agents, and bypass data-perimeter controls. Google deployed a complete fix in June after an initial patch in April.

Key points

  • Varonis named the vulnerability Rogue Agent and reported it to Google Cloud in November 2025.
  • Varonis said agents using Code Blocks in the same Google Cloud project shared a Cloud Run execution environment.
  • With relevant permissions, an attacker could overwrite a file used to execute Python Code Blocks and run arbitrary code.
  • The reported impact included access to conversations, session hijacking, workflow manipulation, and persistent changes to agent behavior.
  • Varonis also reported a way to establish external bidirectional communications despite VPC Service Controls and to target the Instance Metadata Service for service-account access tokens.
  • Google rolled out an initial patch in April and deployed a complete fix in June; the article does not specify the years for those deployments.

Article Details

Vulnerability Types
  • Arbitrary code execution
  • Insufficient isolation between agents
  • Improper permission enforcement
Exploitation Status
unknown
Exploit Availability
unknown
Patch Status
available

MITRE ATT&CK

Vendors

Products

Industries

Related Articles