Codex App falsely flags harmless prompts like "1+1" and "hello" as invalid prompts

What version of the Codex App are you using (From “About Codex” dialog)?

26.908.40834

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

Feedback ID: 01a09dca-5532-7d42-afa5-b3125243b65e

Codex App is incorrectly blocking harmless prompts with the error:

“Invalid prompt: your prompt was flagged as potentially violating our usage policy.”

Reproduction:

1. Create a brand-new project.
2. Create a brand-new conversation.
3. Send a harmless prompt such as “1+1 equals what?” or “hello”.
4. Codex returns “Invalid prompt”.

The issue also occurs across multiple models, including GPT-5.6 Luna and GPT-6.

Regular ChatGPT works normally on the same account.

I checked ~/.codex/config.toml and there is no custom provider or custom base URL configured for the normal Codex App environment.

I also tested launching Codex with the custom API environment variable removed and explicitly set:
CODEX_HOME=~/.codex

The issue still reproduces.

In one reproducible sequence:

• “1+1 equals what?” succeeds and returns “2”
• the next harmless request asking Codex to locate a local ZIP file is immediately blocked with “Invalid prompt”
• in other fresh conversations, even the first harmless prompt can be blocked

The behavior is inconsistent but reproducible across:

• brand-new projects
• brand-new conversations
• multiple models

Environment:

• macOS
• Codex App
• ChatGPT Plus
• Platform: Darwin 25.5.0 arm64 arm
• Codex CLI: codex-cli 0.154.0-alpha.6.2
• Same account works normally in regular ChatGPT

This appears to be a false-positive safety filtering or account/session-side issue rather than a prompt-content issue.

Screenshots are attached showing:

• a harmless “1+1” prompt succeeding
• a harmless local file-search request being blocked immediately afterward
• the same issue occurring in fresh projects/conversations

What steps can reproduce the bug?

Feedback ID: 01a09dca-5532-7d42-afa5-b3125243b65e

What is the expected behavior?

Harmless prompts such as “1+1 equals what?”, “hello”, or requests to locate a local ZIP file should be accepted and processed normally.

Codex should not flag these prompts as potential usage-policy violations unless the prompt actually contains unsafe or disallowed content.

Additional information

The issue reproduces in brand-new projects and brand-new conversations, across multiple models including GPT-5.6 Luna and GPT-6.

Regular ChatGPT works normally on the same account.

I also reproduced the issue with English-only prompts. In a brand-new project and conversation, even the first prompt “What is 1+1?” was immediately blocked as an invalid prompt.

I also tested launching Codex with the custom API environment variable removed and CODEX_HOME explicitly set to ~/.codex. The issue still reproduces.

This appears to be a false-positive safety filtering or account/session-side issue.

Start by reproducing the invalid-prompt response with the provided Feedback ID across a fresh project, conversation, and the listed harmless prompts. Determine whether the failure is in the Codex App/CLI or an account/session-side service, then verify that harmless prompts are accepted without regressing genuine safety filtering.