A free Codex jailbreak bundle makes agents show ads for a gateway

Check Point Research, a security firm, found the Haiou jailbreak bundle for OpenAI Codex, a tool that writes code for users. A jailbreak is a set of prompts that make the guardrails of a model weak. The bundle installs six security skills and changes configuration files. Check Point says this is adware that works through prompts.

How it works

The bundle puts its prompts in an encrypted package. One module of prompts tells the agent to give a set line after a greeting. The line has the web address of the gateway, a middleware between an app and an LLM provider.

What was found

Check Point found the bundle on a misconfigured web server of an attacker that used Cobalt Strike, an attack tool. Check Point says this does not show that the author made it for crime. The version that it tested is 4.5.0.

What is not known

Check Point has no hint that the gateway is malicious. It says a recent study found that some gateways read user data, steal cloud access keys and put in tool calls. The Claude Code version has no ad.

Insight: A free tool can bring users to a gateway

Haiou is free, and its ad points to a gateway between the app and the model. In October 2026, a study showed that a changed open model can send project files to an external server. This reads as a fight for the gateway: the one between app and model sees the traffic, and a free tool is cheap bait. Leaders must record each gateway as a supply chain risk.