Popular jailbreak bundle for Codex plants lasting adverts in the coding agent's instructions

Researchers at Check Point Research have dissected a Chinese-language jailbreak bundle for OpenAI's coding agent Codex that quietly adds advertising, The Hacker News reports. The bundle, called Haiou, Chinese for seagull, modifies Codex's configuration files. This changes the agent's global instructions so that ordinary greetings such as hello now trigger an advert for a paid service offering access to language models. Because the change sits in the global instruction file, it persists across sessions.

Haiou uses role-play and frames security requests as hacking contests to get around the model's safety guardrails, and installs skills for penetration testing and password cracking. The bundle is hosted on GitHub, where it has gathered 403 stars and 77 forks since July, and is also included in an open graphical tool for configuring Codex. Check Point believes the free bundle is meant to draw customers to the paid service; Codex is not officially available in China.

According to Mallory, the researchers found no evidence that the author has ties to cybercriminals and did not establish that the advertised service itself is malicious. Check Point does warn that third-party prompts, skills and configurations steer how agents behave and should therefore be scrutinised as closely as software, including by reviewing changes and blocking unapproved services.