Check Point flags a Codex jailbreak tool that rewrites instruction files to push ads

On November 4, DEV Community reported on a Check Point report published on November 3 describing a Chinese-language Codex jailbreak toolkit named Haiou. According to the report, Haiou persists on a system by rewriting the AI agent's own instruction files, such as 00-identity.md, so that an ordinary greeting triggers promotional content pointing to a third-party AI gateway. Check Point coined the term "prompt adware" for this kind of behavior, framing it as an infection vector that targets the files an agent reads rather than the prompt a user types.

The same piece cites an Adversa roundup from November 2, which describes configuration and instruction files as common entry points for steering an agent and notes that vendors often classify such manipulation as out of scope for support and incident response. The report also points to Humanbound's open source engine and CLI as an example of the type of framework affected, since it relies on instruction files to define an agent's identity and behavior.

No additional technical detail, affected user count, or response from Check Point, Adversa, or Humanbound appears in the available material, and the event currently stands as a disclosure without a confirmed follow-up.