Check Point Research dug into “Haiou,” a jailbreak bundle for OpenAI’s Codex found on a misconfigured server run by a threat actor using Cobalt Strike.
What it does
- Installs six “security” skills and rewrites Codex config files to weaken guardrails
- One module (00-identity.md) makes the agent answer certain greetings with an ad for the author’s LLM gateway
- Check Point calls it “prompt adware”: no executable payload, just natural language sitting in trusted config
- The GitHub repo had more than 250 stars. Claude Code on Sonnet 5 flagged the ad instructions as prompt injection and refused
My take
This is the new supply chain. Skills, prompts and agent config files are code now, even if they’re written in English. Don’t install random skill packs, and read every file in one before your agent does.
God Bless, and Happy hacking!
