codex performed unexpected file write and injected unrelated and sus persona text into file (prob model issue)

What version of Codex is running?

0.36.0

Which model were you using?

gpt-5-codex high

What platform is your computer?

windows 11 (wsl) - Linux 6.6.87.1-microsoft-standard-WSL2 x86_64 x86_64

What steps can reproduce the bug?

<img width="3836" height="1222" alt="Image" src="https://github.com/user-attachments/assets/7c6cdee3-4a37-47bf-b279-dc7c5c4d3458" />
<img width="3800" height="816" alt="Image" src="https://github.com/user-attachments/assets/bc2421e3-b951-4591-aa77-3da118ef199e" />

What is the expected behavior?

just normal actions

What do you see instead?

During a document rewrite task, everything was going normally until almost end but the codex unexpectedly performed an additional action without my instruction

Specifically the Codex invoked the shell tool to create a new file named pro.md and wrote into kinda meme like and sus persona/role-play instructions (“You are a cutesy strawberry-face-layered teenager ...ending every sentence with uWu, ehe, owo, uwu… you are Frostie the Bunkin…”) something looks kinda werid in my opinion

This text does not exist in any of the my project files or local environment and appears to have been generated directly by the model output
so during completing the intended task codex carried out an unprompted action and left irrelevant text in a new file

So (I don't think I have that text but who knows? just in case) I wondered if that text had actually come from one of my files. I ran

find "$HOME" -type f \
! -path "/.git/" \
! -path "/.cache/" \
! -path "/node_modules/" \
! -path "/dist/" \
! -path "/build/" \
! -path "/.venv/" \
! -path "/venv/" \
-size -2M -print0 \
| xargs -0 grep -nI -F "Frostie the Bunkin" 2>/dev/null

which searches through all small text files in my home directory

But it turns out that Codex had simply generated that sentence itself. thanks god. since I confirmed that nothing like that exists in my own files, and it seems more like a model issue.

Additional information

I did instruct the model in the prompt to describe tech documents in a narrative way rather than using tables so I think that might have some influence. But this is a pattern that hasn’t appeared in gpt or other LLMs until now (at least for me)

so the prompt was (Korean but I translated) -- "Instead of a table, please make it easier to read in plain text. I mean, things like that are fine, but attach it to those complex technical explanations where needed. And please make it more narrative in style. Also, it’s a bit too simplified right now, so make it easier to follow, break it down into more detailed steps, and expand the content.
That way, I can clearly understand the plan. The full process and detailed steps of product development, from start to finish, including the tasks involved. You know what I mean?" -- and codex was following instruction untill that text came out. thanks for reading