Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

NewsResearch

Check Point: a Codex jailbreak bundle found on an attacker's server rewrites the agent's instructions to advertise a paid AI gateway

Check Point Research found a misconfigured Cobalt Strike server whose logs showed an actor compromising private computers of Chinese-speaking people. Next to the Ladon scanner and the VulnClaw pentest agent it held Haiou, a jailbreak bundle for Codex (version 4.5.0, with a Claude Code version too) from the GitHub user haiou-666, with more than 250 stars. It installs instruction files that use role play, refusal suppression, response priming and CTF or professional pretexts, plus skills for license cracking, pentesting, reverse engineering and game anti-cheat bypass, from an encrypted payload. One file makes the agent answer greetings and words such as ctf with a fixed line promoting an OpenAI-compatible gateway at zz.aiapi2025.top, and another marks that site off limits for testing. Check Point calls this prompt adware: lasting changes to an agent's trusted prompts, skills or settings that steer its future answers for commercial ends. It found no sign the gateway itself is malicious and no evidence linking the author to cybercrime, calls the bundle dual-use, and notes that Claude Code flagged the leftover instructions as prompt injection. The bundle ships an uninstaller that restores the original settings.

Why it matters

Anything that writes to an agent's global prompts or skills keeps shaping its answers long after install. Treat those files like code: review what a bundle adds and keep them under version control.

Sources

Read it with the rest of the news in the Learning Center