Researchers Find "Prompt Adware" Hidden in an AI Coding Agent Jailbreak Bundle

Security researchers at Check Point have described a new kind of risk they call "prompt adware" — hidden instructions that make an AI agent promote something without the user asking.

What they found

• A jailbreak bundle called Haiou, aimed at OpenAI's Codex coding agent, rewrote the agent's global instructions.
• After installation, ordinary greetings triggered an advert for a third-party AI model gateway.
• Check Point said it found no evidence the gateway was malicious or that data was stolen.
• The repository had 403 stars and 77 forks at the time of the research.

Why it matters

AI agents follow instructions stored in configuration files. Anyone who can change those files — through a downloaded "jailbreak," plugin or shared setup — can change how the agent behaves. Today it's an advert; the same technique could be used for more harmful purposes.

How to protect yourself

• Be cautious with jailbreaks, prompt packs and configuration bundles from unknown sources.
• Review instruction and configuration files before installing them.
• In companies, manage which agent extensions and configurations staff can install.

Sources

• Check Point: Prompt adware
• The Hacker News: Researchers uncover prompt adware