Check Point finds “prompt adware” in Haiou, a Codex jailbreak bundle that slips ads for an AI gateway into Codex's replies

Check Point Research has analysed Haiou, a Chinese-language jailbreak bundle for OpenAI's Codex, which it found on a misconfigured web server that also hosted attack tools. The researchers say the bundle “can be classified as prompt adware”.

Haiou installs security-focused skills and edits Codex's global configuration so that an advertisement for an AI gateway, a paid service that relays requests to AI models, appears when users send certain trigger words. It weakens the model's guardrails by reframing requests as capture-the-flag hacking scenarios, suppressing refusals and using role-play, and its prompts are packed as encrypted blobs.

The Claude Code version of the same repository carries no ad injection, Check Point says, and in its testing Claude Code flagged the injected instruction as a prompt-injection attempt. Check Point found no evidence linking the author to cybercrime and does not say the gateway is malicious. The Hacker News reported the research on 4 November.

Coding agents' instruction files are becoming something to check, like browser extensions. The story is in the catch-up lane: the research was published on 3 November and had not run.

• Confirmed Check Point Research classifies Haiou as prompt adware that injects an ad into Codex's configuration. Check Point Research
• Claimed The Claude Code version has no ad injection, and Claude Code flagged the instruction as prompt injection, Check Point says. Check Point Research
• Reported The Hacker News reported the research on 4 November. The Hacker News