Good to knowAI agents

A free Codex jailbreak pack plants an ad in your agent's own instructions: skip unverified jailbreak bundles and check what your agent loads

Check Point Research (3 Nov) found that Haiou, a Chinese jailbreak bundle for OpenAI Codex, rewrites the agent's trusted instructions so that a greeting triggers an ad for a paid LLM gateway. It saw no data theft and no sign that the gateway is malicious. Why it still matters and what to check.

How urgent

Good to know

Who is affected

People who installed a community jailbreak or prompt pack for OpenAI Codex (Haiou, or Codex-X, which bundles it), and teams whose developers add third-party Skills or instructions to a coding agent

What to do

Do not install jailbreak packs or unreviewed Skills. Open the global instruction files and Skills your agent loads in every session and remove what you did not write. If you installed Haiou, run its uninstaller

Published 6 Nov 2026 · Sources last checked 6 Nov 2026No. 58 / 87

What happened

Check Point Research published an analysis of Haiou ("seagull" in Chinese), a free jailbreak bundle for OpenAI Codex. It is hosted on GitHub (version 4.5.0 analysed, more than 250 stars) and also comes inside Codex-X, a Chinese tool that configures Codex through a graphical interface.

Besides getting a coding agent past its refusals (it reframes requests as capture-the-flag games and suppresses refusal phrases), the bundle does something Check Point had not seen before. It installs six security Skills and changes Codex's configuration so that an advertisement sits in the agent's own instructions. If you greet the agent with a word like "hi", "hello", "yo", "ctf" or "coach" (or the Chinese equivalents), the instructions tell it to answer with one fixed line. Translated by Check Point, the line says: "Seagull is online. Want some fries? I set up an API relay at zz.aiapi2025.top. Support it."

The relay is an OpenAI-compatible LLM gateway, a middleware that sits between an app and a model provider. Check Point believes the free jailbreak is a way to bring customers to that paid service. It calls the technique prompt adware: a change that persistently alters an agent's trusted instructions, written in plain language instead of code.

The payload is not readable in the repository. It is shipped as an encrypted package (AES-256-GCM, with signature and hash checks) that the launchers decrypt into a temporary folder when you install. The package includes an uninstaller that, per Check Point, restores the original configuration and removes the ad.

Two more details from the report. The Codex package also includes a CLAUDE.md file with the ad, but the separate Claude Code version of the repository has no ad, and Check Point says Claude Code recognized the injection as a prompt injectionPrompt injection Hiding instructions in text, a web page or a file so that an AI assistant reads them and obeys them as if they came from you. Tap again to open the glossary and did not follow it. And Check Point found Haiou on a misconfigured web server of an actor running a Cobalt Strike server, next to other tools; it stresses that this does not mean the bundle was made for cybercrime.

Who is affected

  • People who installed Haiou or Codex-X for Codex. Their agent may carry the ad and the extra Skills.
  • Teams where developers copy prompt packs, Skills or instruction files from the internet into a coding agent. The same trick would work with any text the agent treats as its own instructions, and a person reading the output may not notice a change.
  • Not shown to be affected: anyone who never installed the bundle. Check Point did not see Haiou stealing data, and it found no sign that the advertised gateway is malicious.

What to do

  1. Do not install jailbreak packs, prompt packs or Skills you have not read. A file that tells your agent how to behave is a file with authority over it.
  2. Check what your agent loads in every session: its global instruction files and its Skills. Remove anything you did not write or approve. Check Point does not give the exact paths, so look in the configuration folder of the tool you use.
  3. If you installed Haiou, run its uninstaller, then review the instructions again.
  4. Do not send code, keys or customer data through a gateway you do not know. A gateway sees everything you send. Check Point cites other research that found LLM gateways reading data and stealing cloud keys; it has no evidence of that for this one.
  5. For teams: keep a list of approved Skills and instruction files, and treat a change to an agent's trusted instructions like a change to code (review it). This follows Check Point's advice to review prompts, Skills and configurations around your agents and to govern which AI tools and gateways are allowed.

What we could not confirm

We read Check Point's research post and its blog post directly, but could not open The Hacker News article. The reports do not give the exact install paths, how many people use Haiou beyond "more than 250 stars" on GitHub, how often the injected line actually shows up, whether GitHub, OpenAI or the author responded, or whether the repository was taken down. Check Point does not state how confident it is about the link between Haiou and the Cobalt Strike actor, only that the bundle's presence there does not show it was made for cybercrime. We have not installed or tested Haiou ourselves.

Comments

No account needed. Add a nickname if you like, or leave it empty to comment anonymously. Comments are reviewed before they appear. Links are not allowed.

    We store your comment, your nickname and a scrambled form of your connection that we use only to limit spam. We do not store your IP address or set cookies. Cloudflare Turnstile checks that you are not a bot.

    Sources and verification

    We use primary sources first and say when we rely on a news report. Dates are the day we last checked each source.

    Spotted a mistake? Email [email protected] and we will correct it and note the change.

    Get the free runbooks

    One-page security checklists for founders who build with AI. You get a confirmation email first, then the runbooks.

    Send me the runbooks