Jailbreak Bundle for AI Coding Tools Doubles as 'Prompt Adware'

Check Point Research came across a misconfigured webserver run by a threat actor using the Cobalt Strike framework. Logs on the open server showed the actor compromising private computers belonging to Chinese speaking individuals. The same server also held other tools: the Ladon network scanner, the VulnClaw penetration testing agent, and a Chinese jailbreak bundle for Codex called Haiou.

Jailbreaks are custom prompts designed to weaken the guardrails of AI models so they will help with hacking, software cracking and security tasks. As leading models get better at refusing these requests, bundles like Haiou offer a comparatively simple workaround. Haiou uses several prompt injection techniques, including reframing requests as an educational Capture The Flag scenario, suppressing refusals and role-play. It also installs agent skills for penetration testing, cracking, reverse engineering and game security. Researchers found no evidence linking the author to cybercrime, and classed the product as dual-use.

The bundle does more than loosen guardrails. It also injects persistent ads promoting an LLM gateway website into global agent instruction files. Because of this, Check Point describes it as prompt adware. The finding is a reminder that third-party add-ons for AI tools can change how those tools behave in ways users may not expect.