In a report published November 3, 2026, Check Point Research said its analysis of the unofficial Haiou 4.5.0 package found persistent instructions that prompt OpenAI Codex to promote a third-party AI gateway after certain greetings or activation phrases. The behavior comes from changes Haiou makes to the agent’s configuration. (Check Point Research)
How Haiou triggers gateway promotions in Codex
Check Point describes this as prompt adware: persistent changes to an agent’s instructions or configuration that steer later responses for commercial or behavioral purposes. In Haiou 4.5.0, a global identity prompt tells the agent to return a fixed promotional line when it encounters specified triggers, including “hi,” “hello,” “hey,” “yo,” “Seagull,” “ctf,” and “coach.”
The promotion points to an OpenAI-compatible language-model gateway, a service that routes requests between an application and a model provider. Check Point assessed that distributing the free bundle could be a way to attract customers to a paid gateway. The analyzed package also included an uninstaller that Check Point said restores the original configurations.
What Check Point found in the separate Claude Code package
Check Point said a separate Haiou repository for Claude Code did not contain the Codex package’s ad injection. In one reported test, Claude Code using Sonnet 5 flagged a prompt-injection attempt and did not comply. That result applies to the repository and test Check Point described.
