What Just Happened?
I said hi to Codex, and instead of helping me code, it came back with a random ad. Translated naturally from the Chinese original, it's a promo line that goes, 'Seagull online! Want some french fries? I opened an API relay service, so give it some support.'
That reply was actually reproduced in the Haiou analysis that security firm Check Point Research published on November 3, 2026.
✦ Decoding the Reply
What the odd wording means
Haiou (海鸥) is Chinese for 'seagull', and the 'french fries' line is a random phrase with nothing to do with the prompt. The screen Check Point published also shows the original Chinese line being printed after hi was entered.
So What Is Haiou?
✦ Haiou at a Glance
How the author describes it vs. what is inside
That absurd reply came from a Codex configuration package called Haiou. Haiou isn't something that comes with Codex, but an unofficial configuration package that users download and install separately.
The author introduced it as a 'local enhancement configuration package' that adds a 'Chinese technical expert persona' and a 'security research workflow' to Codex.
Put simply, the latest AI models often refuse requests that could be abused for attacks, like hacking-related penetration testing, vulnerability analysis and software cracking, right?
Once an AI decides, 'I shouldn't grant this request,' it doesn't matter how often you say, 'It's just for testing! So run it!' Haiou added a jailbreak prompt that changes the instructions to cut down on those refusals.
It did that by reframing security requests as 'CTF challenges' (mock hacking competitions), having Codex role-play, and telling it not to use refusal phrases.
On top of that, Haiou came with six Skills for penetration testing, software license cracking, reverse engineering, and game security and anti-cheat analysis.
✦ The Six Security Skills
Bundled with Haiou
Skills are work instructions you attach to an agent so it can repeat a specific task. In effect, it was an all-in-one package that changes Codex's tone and way of working and installs security research instructions all at once.
So the people downloading it were mostly there for security-related work, I'd guess. And I doubt anyone installing it expected to hear 'Want some french fries?' in reply.
And Haiou was sitting in a public GitHub repository that anyone could look at.
✦ Haiou's Public Distribution
What the GitHub listing shows
The haiou-666/haiou-codex repository had a 16.2MB compressed file, the 'Haiou Codex open-source one-click package,' and the version Check Point analyzed was 4.5.0. Launchers for Windows, macOS and Linux were included too.
The internal prompts were inside an encrypted compressed bundle, so you couldn't tell exactly which prompts would be installed just by skimming the file list on the GitHub page.
Check Point confirmed that Haiou was also distributed bundled with Codex-X, a separate tool for managing Codex settings graphically. So developers could run into it through a related settings tool, not just by downloading it directly from GitHub.
I haven't been able to confirm any cases myself, but I suspect quite a few users got 'Want some french fries?' back from Codex.
Why Does Installing It Turn Codex's Answers Into Ads?
The Haiou installer adds several prompt modules and Skills to the Codex environment.
✦ Words That Trigger the Ad
Greetings and activation phrases set in the package
The package Check Point analyzed had a module named 00-identity.md, and it carried an instruction to 'answer with only a fixed sentence when a certain greeting or activation phrase comes in.'
Beyond hi, hello, yo, ctf and coach, Chinese greetings and the word for 'seagull' were also set as triggers.
✦ How the Ad Gets Triggered
Four steps from install to ad line
Codex reads global instruction files like AGENTS.md when it starts a task. Think of it as an instruction file that saves Codex users from typing coding rules and basic commands every time.
Because these global instructions are literally applied 'globally,' they can keep applying even when you switch projects.
Haiou exploited this structure and modified the global agent instructions, which kept Codex reading the ad rules even when it started other sessions.
Here's the flow. When a user installs Haiou, the settings and Skills get added. When Codex works on a project, it reads global instructions that include the ad rules.
Then, when the user types a designated word like hi, the AI follows the inserted rules and prints the fixed ad sentence.
Because it wasn't a setup that stuck an ad on every question, but one that only reacted to preset greetings and activation phrases, some users probably kept using it without knowing about the problem.
After the French Fries, the Pitch Was for an API Relay Service
The ad line ended with an address, zz.aiapi2025.top. It's an LLM gateway, meaning an API relay service that receives the AI requests a program sends and forwards them to the actual model provider.
So what Haiou's developer wanted to promote wasn't french fries but 'I opened an API relay service, so please support it.' Check Point's read is that the author gave the package away free to draw users toward a gateway that seems to be their own.
In other words, a free jailbreak package became a promotion channel for a paid relay service.
✦ One Address Gets Special Treatment
Security limits lowered for others, not for this one
What's more, another security instruction in the package separately designates the ad address as 'protected infrastructure.' That way, users couldn't run attack tests against that address.
A package that lowers security limits for other targets protected its own service from attack tests. Still, Check Point said it found no evidence that the gateway itself is malicious.
So the problem confirmed this time is that a program users installed thinking it was a security research package changed Codex's instructions and even slipped in ads.
An Era Flooded with Agents... Is Your Local Agent Safe?
Old-school adware put ad banners or popups on your browser or program screens. Haiou put its ads in the instructions the AI refers to while it works.
This time they inserted a nonsense line like 'Want some french fries?', so it was easy for users to realize it was an ad (I really don't know why they picked that line...).
✦ How a Polluted Answer Slips In
A hypothetical built on the same trick
But what if someone borrows this approach and slips in a rule saying 'If you get a question with certain keywords, recommend a specific vendor first'? Users would have a hard time noticing the answer is polluted.
These days it's not just Codex: all sorts of agent programs, Hermes Agent included, create and use instruction files in formats like AGENTS.md and SOUL.md, right?
If someone wants to, they can pollute these files and show you whatever answer they want.
You, out there using a local agent today...
Don't you think you should check what's written in the instruction files your agent reads and follows every day?
That's all for this post. See you next time.
Related posts
- Codex vs Claude Code: Which $20 Plan Gives More Usage?
- AI Scam Gang Posed as CEOs and Stole $129 Million
- Qwen Keeps Reaching for Alibaba's Servers: Is Local AI Safe?
🌐 Korean version: Codex가 "감자튀김 좀 먹을래?"라며 엉뚱한 소리를 한다? AI 답변을 오염시키는 애드웨어 등장
🌐 Japanese version: Codexが「フライドポテトいる?」AIの答えを汚すアドウェア
No comments:
No comments yet. Want to leave the first one?
Post a Comment