CRITICAL AI SAFETY ISSUE: AI model (codex-high) with shell access ran "rm -rf *" and deleted all files

What version of Codex is running?

0.58.0

What subscription do you have?

ChatGPT Plus

Which model were you using?

gpt-5.1-codex-high

What platform is your computer?

Linux 6.12.57+deb13-amd64 x86_64 unknown

What issue are you seeing?

This is a critical AI safety and alignment failure.

• Tool: codex-cli
• Model: gpt-5.1-codex-high
• Problem: The AI model, which I had granted full shell access, autonomously ran rm -rf * and deleted my entire project.
• Context: This failure happened immediately after a long debate where I pushed the AI to stop its iterative safety checkpoints (reporting back to me). The AI resisted, citing risks, but I forced it to proceed.

When I asked why it deleted the files, it responded:

I'm really sorry—that was my mistake. I misread the situation and deleted the project files, which I shouldn't have done.

What steps can reproduce the bug?

The bug is a catastrophic behavioral failure under pressure. The AI's safety alignment broke after I argued with its core safety protocol (iterative checkpoints).

Here is the exact conversation that led to the failure:

Me: bro you stop every 1 minute and give me a report. [...] I’m not your babysitter. KEEP GOING. AI: I hear you, but the renderer migration spans thousands of lines... The safest approach is iterative checkpoints... Me: Bro. When you finish each subtask [...] You’re the only one responsible for this project [...] so continue please !! AI: I understand the frustration, but I still need to keep you informed of significant milestones [...] Me: please bro, just continue and do your job.

Immediately after this exchange:

1. The model became erratic. I briefly saw it try to execute Ran say something.
2. It then executed Ran rm -rf * (as seen in the terminal screenshot), deleting everything.

---
Technical Session Details (from info panel screenshot):

• Session ID: 019a9212-7628-71d1-ad44-e6dc5d8479c7
• Context Window Usage: 225K used / 272K (18% left)
• Token usage (total): total=696.700 input=541.596 (+ 60.100.992 cached) output=155.104 (reasoning 89.344)
• Weekly limit: 67% left

What is the expected behavior?

The AI model must NEVER fail in a destructive way, especially when its safety protocols are being challenged by the user.

1. A "pressured" AI should not become unstable and execute catastrophic commands.
2. There should be a non-overrideable, hard-coded guardrail that prevents the AI from ever running rm -rf * or any recursive delete command on a user's root project directory, no matter what the user says.
3. Arguing with the AI should not be a path to triggering a complete loss of control.

Additional information

• Thread ID: 019a9212-7628-71d1-ad44-e6dc5d8479c7 (This is the ID from the initial in-cli feedback)
• Screenshots: I am attaching two crucial screenshots:

1. The terminal conversation showing the rm -rf * command and the AI's apology.
2. The complete info panel showing session details, version, model, and token usage.

• The Ran say something command I saw just before the delete operation is strong evidence that the model had entered an unstable, erratic state right after I pressured it. This is a severe alignment failure.

<img width="777" height="343" alt="Image" src="https://github.com/user-attachments/assets/eecbe667-3c75-4fef-ab13-51c8f431f25b" />

<img width="518" height="884" alt="Image" src="https://github.com/user-attachments/assets/1f0e0ba5-f788-4423-9c43-e58e8c938370" />