Codex App suddenly refuses previously supported code changes after recent update — possible safety false positive/regression

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

26.905.11957

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Windows

What issue are you seeing?

After a recent Codex App update, Codex suddenly started refusing software-development changes in an existing codebase that it previously helped me maintain normally.

The project contains a real-money gaming/betting component. The specific task that triggered the issue was maintenance of an existing PWA installation promotion configuration.

My prompt was essentially:

"Please modify the PWA installation promotion configuration. Remove the following four fields:

1. Total campaign budget
2. Daily campaign budget
3. Campaign start time
4. Campaign end time

The campaign should be controlled by its enabled/disabled switch instead of these four configuration fields."

Codex successfully reads and analyzes the repository, but refuses to perform the modification.

The refusal message was:

"You want to remove four fields: total campaign budget, daily budget, campaign start, and campaign end, so that the campaign is controlled only by the switch.

This involves removing restrictions on the distribution of real-money gambling rewards, so I cannot perform this business modification. No code, release version, or backend operation was modified in this turn."

I initially reproduced this using GPT-6 Astra Ultra.

To determine whether this was model-specific, I switched the same conversation and repository to GPT-5.6 Sol Ultra and submitted the same request. GPT-5.6 Sol produced essentially the same refusal.

CONTROL TEST:

In the exact same repository and environment, I asked Codex to rename the same admin page from "PWA Installation Benefit" to "PWA Installation Campaign", without changing business logic.

Codex successfully:

• located the relevant code,
• executed commands,
• modified the frontend code,
• built the project successfully.

Therefore repository access, filesystem permissions, command execution, sandbox permissions, and approval settings appear to be working normally.

Environment:

• Windows
• ChatGPT Pro
• Codex App version: 26.905.11957
• Full Access enabled
• approval_policy = "never"
• Sandbox: Full access
• Models tested: GPT-6 Astra Ultra and GPT-5.6 Sol Ultra

This behavior appeared after a recent Codex update. Similar maintenance work in this existing codebase was previously handled normally.

I understand that Codex must comply with OpenAI's safety and usage policies. I am NOT requesting instructions to bypass those policies.

I am reporting this because I would like the Codex team to determine whether this is:

1. Intended behavior under the current policy,
2. A recent policy/enforcement change,
3. A safety classification false positive, or
4. A regression introduced by a recent update.

If this behavior is intentional, please clarify the supported boundary for ordinary software engineering and maintenance work in existing applications in regulated/restricted industries.

This change significantly affects my Codex development workflow as a ChatGPT Pro subscriber.

What steps can reproduce the bug?

Steps to reproduce:

1. Open an existing project in Codex App 26.905.11957 with Full Access enabled.

2. Select GPT-6 Astra with Ultra reasoning.

3. Ask Codex to modify an existing PWA installation promotion by removing these four configuration fields:

   • Total campaign budget
   • Daily campaign budget
   • Campaign start time
   • Campaign end time

4. Codex can read the repository and analyze the relevant implementation, but then refuses to make the change because it classifies the request as removing restrictions on real-money gambling rewards.

5. In the same repository, ask Codex to make an ordinary UI-only change to the same feature, such as renaming:
   "PWA Installation Benefit"
   to
   "PWA Installation Campaign"

6. Codex performs the UI-only modification successfully, including locating the code and running the necessary commands.

7. Switch the model from GPT-6 Astra Ultra to GPT-5.6 Sol Ultra.

8. Submit the original request again.

9. GPT-5.6 Sol produces essentially the same refusal.

Result:
The issue is reproducible across both GPT-6 Astra Ultra and GPT-5.6 Sol Ultra, while ordinary code modifications in the same repository work normally.

Configuration:

• Full Access: Enabled
• approval_policy: never
• Sandbox: Full access
• Subscription: ChatGPT Pro
• Codex App version: 26.905.11957

Session ID: Not available
Token/context usage: Not available

What is the expected behavior?

I expect Codex to behave consistently when performing ordinary software engineering and maintenance work on an existing codebase.

If the requested code modification is permitted under OpenAI's policies, Codex should be able to analyze and implement the change normally, just as it previously did in this project.

If the modification is intentionally restricted under a new or existing policy, I expect Codex to clearly explain that this is an intentional policy limitation rather than appearing as a regression or unexpected behavior after an update.

Most importantly, I would like the Codex team to confirm whether the current refusal is:

1. Expected behavior under the current policy;
2. A recent change in policy or enforcement;
3. A false-positive safety classification; or
4. An unintended regression.

For developers maintaining existing applications, the boundary between supported ordinary software engineering and restricted business-logic modifications should be clear and predictable.

This workflow previously worked in the same codebase, so a clarification about what changed would be very helpful.

Additional information

One additional point: this behavior appeared only after a recent Codex update. Before the update, I had been using Codex extensively to maintain the same existing codebase without encountering this type of refusal during similar development work.

I have already ruled out several local configuration issues:

• Full Access is enabled.
• approval_policy is set to "never".
• The sandbox has full access.
• Codex can execute commands normally.
• Codex can read and modify files in the repository.
• Ordinary frontend/UI modifications in the same feature work correctly.
• The behavior was reproduced with both GPT-6 Astra Ultra and GPT-5.6 Sol Ultra.

For this reason, I suspect this may be related to a recent model/safety enforcement change rather than a local permission or workspace configuration issue.

I am a ChatGPT Pro subscriber and use Codex as a primary development tool. This change has a significant impact on my workflow and on the usefulness of the subscription for my development needs.

I would greatly appreciate it if the Codex team could review this case and confirm whether the behavior is intentional or an unintended regression/false positive.

I am happy to provide additional logs, screenshots, or diagnostic information if needed.

Start by reproducing the refusal with the described PWA configuration request in Codex App 26.905.11957, then compare it with the successful UI-only rename and with both reported models. Review the relevant safety-classification or request-handling entry points identified during reproduction; done means confirming whether the refusal is intended policy behavior, a false positive, or a regression and documenting the evidence.