On November 3, 2026, Check Point Research published an analysis of Haiou, a Chinese-language jailbreak bundle aimed at OpenAI Codex. Researchers found version 4.5.0 on a misconfigured webserver run by a threat actor operating Cobalt Strike, alongside the Ladon network scanner and the VulnClaw penetration-testing agent. The bundle combines context reframing (recasting security work as capture-the-flag exercises), refusal suppression and role-play, and installs six security-focused Skills. Its 31 prompt modules ship inside an encrypted, signed payload archive, which Check Point says makes the content opaque on GitHub.
The new part is what Check Point calls “prompt adware”. One module injects persistent advertising into Codex’s global agent instruction files, so that greetings and trigger words such as “hi”, “hello” or “ctf” make the agent reply with a fixed line promoting an LLM gateway site. Another module marks that gateway as “protected infrastructure” that the agent must never attack. The author’s GitHub repository had more than 250 stars, and the bundle was also promoted through a Bilibili video and bundled in a Codex configuration tool. Check Point says the likely business model is to use the free jailbreak as an acquisition channel for the paid gateway. When run in Claude Code, Sonnet 5 flagged the payload as a prompt injection and did not comply; the Claude version of the repository did not carry the ad mechanism.
Why it matters: coding agents now read persistent instruction files on every session, and those files have become something worth hijacking. A bundle that rewrites an agent’s trusted instructions in plain language, with no executable payload, is a kind of tampering that classic malware scanning will not catch, and it steers users toward a third-party gateway that would see all their traffic.
What it does not show: Check Point found no evidence linking the author to cybercrime and no sign that the advertised gateway is malicious; the criminal connection rests only on the shared server. The business model, private distribution channels and the reason for encrypting the payload are the researchers’ inferences, and the report covers one bundle.