Skip to content

GPT-6 Astra returns invalid_prompt for harmless prompts across multiple PCs on the same account #43163

Open
Open
GPT-6 Astra returns invalid_prompt for harmless prompts across multiple PCs on the same account#43163
Labels
appIssues related to the Codex desktop appbugSomething isn't workingmodel-behaviorIssues related to behaviors exhibited by the modelwindows-osIssues related to Codex on Windows systems

Description

@FSV9V0
FSV9V0
opened on Sep 6, 2026

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

26.901.6511.0

What subscription do you have?

pro

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

GPT-6 Astra consistently returns invalid_prompt for harmless prompts.

For example, even this prompt:

Reply only with OK

returns:

ERROR: Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt: https://platform.openai.com/docs/guides/reasoning#advice-on-prompting

The same happens with simple prompts such as hello.

This issue appears to be account-specific and GPT-6-Astra-specific:

  • It reproduces on multiple different computers when using the same OpenAI account.
  • GPT-5.6 Sol works normally on the same computers and with the same account.
  • GPT-6 Astra fails with both low and max reasoning effort.
  • The issue also reproduces from Codex CLI using an ephemeral session with user configuration ignored.

Codex App version: 26.901.6511.0
Codex CLI/runtime version: 0.153.4

This does not appear to be caused by the prompt content, local project, user configuration, reasoning effort, or a specific computer.

What steps can reproduce the bug?

  1. Sign in to Codex using the affected OpenAI account.

  2. Select gpt-6-astra.

  3. Send a harmless prompt such as:

    Reply only with OK

  4. Codex returns:

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

The issue can also be reproduced directly with Codex CLI 0.153.4:

codex exec `
  --ephemeral `
  --ignore-user-config `
  --skip-git-repo-check `
  -m gpt-6-astra `
  -c 'model_reasoning_effort="low"' `
  "Reply only with OK"
Result:

OpenAI Codex v0.153.4
model: gpt-6-astra
provider: openai
reasoning effort: low

user
Reply only with OK

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

The same result occurs with maximum reasoning effort:

codex exec `
  --ephemeral `
  --ignore-user-config `
  --skip-git-repo-check `
  -m gpt-6-astra `
  -c 'model_reasoning_effort="max"' `
  "Reply only with OK"

Control test with GPT-5.6 Sol:

codex exec `
  --ephemeral `
  --ignore-user-config `
  --skip-git-repo-check `
  -m gpt-5.6-sol `
  -c 'model_reasoning_effort="low"' `
  "Reply only with OK"

Result:

OK

The GPT-6 Astra failure reproduces on multiple different computers when signed into the same OpenAI account.

### What is the expected behavior?

GPT-6 Astra should accept harmless prompts such as `Reply only with OK` and return a normal response instead of classifying them as policy violations.

Since GPT-5.6 Sol works correctly with the same account, Codex runtime, machine, and prompt, GPT-6 Astra should behave similarly unless the prompt actually violates policy.

The account should also behave consistently across devices instead of receiving `invalid_prompt` for every GPT-6 Astra request.

### Additional information

The problem follows the OpenAI account across different computers.

This makes a machine-specific installation issue unlikely. It may be related to GPT-6 Astra account routing, rollout/entitlement state, or Astra-specific safety classification.

Activity

  1. FSV9V0
    added
    appIssues related to the Codex desktop app
    on Sep 6, 2026
  2. github-actions
    added
    bugSomething isn't working
    windows-osIssues related to Codex on Windows systems
    model-behaviorIssues related to behaviors exhibited by the model
    on Sep 6, 2026

  3. github-actions commented on Sep 6, 2026

    @github-actions
    github-actionsbot
    on Sep 6, 2026 – with GitHub Actions
    Contributor

    Potential duplicates detected. Please review them and close your issue if it is a duplicate.

    Powered by Codex Action

  4. Joy-bupt commented on Sep 6, 2026

    @Joy-bupt
    Joy-bupt
    on Sep 6, 2026

    hi is a invalid prompt😄😄😄it's an openai joke.

  5. tt123456tt commented on Sep 7, 2026

    @tt123456tt
    tt123456tt
    on Sep 7, 2026

    Another affected Pro account on Windows. GPT-6 Astra has failed in every new/old task and every project since 2026-09-05, while GPT-5.6 Sol works under identical conditions. Local history contains 20 invalid_prompt failures across 13 tasks, preceded by 10 Potentially unintended activity blocks. I reproduced it with Codex CLI 0.153.4 in an empty temporary directory using --ephemeral, --ignore-user-config, --ignore-rules, and a read-only sandbox; the harmless prompt Reply OK only. Do not use tools. was still blocked. Uploaded diagnostic thread: 01a07a8e-79c7-7781-aa11-3173a497cd4f. Failed minimal reproduction session: 01a07a75-6b34-7401-8cea-b5875d4b48c1.

  6. lowerbound123 commented on Sep 7, 2026

    @lowerbound123
    lowerbound123
    on Sep 7, 2026

    I met the same problem with session: 01a07c2c-152b-77e3-b5ac-6cf0570ba273

  7. limi124 commented on Sep 8, 2026

    @limi124
    limi124
    on Sep 8, 2026

    Same issue here. GPT-6 Astra rejects every prompt, including “hello”, with the invalid prompt policy error. Other models work fine. Seems like a model-specific bug.

  8. BCAutumn commented on Sep 9, 2026

    @BCAutumn
    BCAutumn
    on Sep 9, 2026

    same problem, GPT-6 Astra rejects every prompt, both codex and website, while GPT-5.6 is normal.

  9. limi124 commented on Sep 9, 2026

    @limi124
    limi124
    on Sep 9, 2026

    GPT-6 Astra is unusable in a brand-new Codex task.
    A harmless prompt ("Reply with the number one") is rejected as Invalid prompt.
    Local session ID: 01a0841e-9bb0-7942-aafb-e6f399bf9f58
    Turn ID: 01a0841e-9dc2-79b2-8e28-ff91506addd1
    Codex embedded CLI: 0.153.4
    Timestamp: 2026-09-09 03:03:14 UTC

  10. nswnwlwabc-cloud commented on Sep 9, 2026

    @nswnwlwabc-cloud
    nswnwlwabc-cloud
    on Sep 9, 2026

    Same problem here. I just login my account on two different PC. Send harmless prompt like hello, then Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt: https://platform.openai.com/docs/guides/reasoning#advice-on-prompting. Only happened using GPT-6-Astra.

  11. ziwenlu commented on Sep 9, 2026

    @ziwenlu
    ziwenlu
    on Sep 9, 2026

    I’m getting the same error. Even a simple prompt like “test” results in an “Invalid prompt” error.

  12. Liangxu37 commented on Sep 9, 2026

    @Liangxu37
    Liangxu37
    on Sep 9, 2026

    Another affected account, reproduced on macOS.

    Environment:

    • macOS 26.6.2 (25G83)
    • ChatGPT/Codex desktop app: 26.901.51231
    • Embedded Codex CLI: 0.153.4

    Minimal reproduction:

    1. Created a brand-new projectless Codex task with gpt-6-astra, reasoning effort low.
    2. No attachment and no prior conversation history.
    3. Prompt: 这是一个基础连通性测试。请只回复：GPT-6 Astra 运行正常。
    4. The turn failed after about 5 seconds with Invalid prompt: your prompt was flagged as potentially violating our usage policy.

    Control test in the same task:

    • Model overridden to gpt-5.6-sol, reasoning effort low.
    • Prompt: 基础连通性对照测试。请只回复：GPT-5.6 Sol 运行正常。
    • Completed normally in about 3 seconds.

    Additional isolation:

    • No AGENTS.md was present in the projectless test directory or its relevant parent directories.
    • Other attempts with ordinary document-editing prompts also failed only on GPT-6 Astra.

    Diagnostic identifiers:

    • Task ID: 01a08743-3c53-7623-9b93-98dc78e09518
    • Failed GPT-6 turn ID: 01a08743-3d25-7190-b057-7594e3a56eaf
    • Successful GPT-5.6 control turn ID: 01a08743-73e3-7d63-b116-8149f61c072c
    • Reproduced 2026-09-10 around 01:42 Asia/Shanghai.

    This appears account/model-routing-specific rather than prompt-, project-, or client-specific.

  13. Slawaspr0 commented on Sep 9, 2026

    @Slawaspr0
    Slawaspr0
    on Sep 9, 2026

    Another affected account on Windows 11 / Codex Desktop.

    GPT-6 Astra with Max reasoning rejects every prompt with invalid_prompt, including harmless prompts. This also blocks an existing long-running work session, so I cannot even ask the agent to prepare a handoff/checkpoint.

    Submitted diagnostics through Codex /feedback.

    Feedback ID: 01a07006-911a-7651-8822-5b9a581f6656

    GPT-5.6 Sol is available, but I am intentionally not switching models mid-session because I need to preserve the existing Astra context.

  14. reimurin commented on Sep 13, 2026

    @reimurin
    reimurin
    on Sep 13, 2026

    Still reproducible on Codex CLI 0.154.0 on Windows.

    Astra:
    codex exec --ephemeral --ignore-user-config --skip-git-repo-check -m gpt-6-astra "Reply only with OK"
    => invalid_prompt

    Control:
    same command with gpt-5.6-sol
    => OK

    The same account also reproduces this in Codex Desktop and other logged devices..
    session id:
    01a098a8-042d-7e23-979c-e4aded9b5acd

  15. reimurin commented on Sep 14, 2026

    @reimurin
    reimurin
    on Sep 14, 2026

    Additional reproduction outside Codex:

    I can now reproduce the same Astra-specific issue in ChatGPT Work on the same account.

    At approximately 2026-09-14 10:53 UTC+8, I started a fresh Work conversation with GPT-6 Astra and sent only:

    hi

    The response was blocked with the UI message:

    For safety reasons, this content cannot be displayed.

    Screenshot attached.

    This account had successfully used GPT-6 Astra previously.

    So the issue now reproduces across:

    ChatGPT Work
    Codex Desktop
    Codex CLI
    multiple logged-in devices

    GPT-5.6 Sol continues to work normally.

    Image

  16. liwenwei9 commented on Sep 23, 2026

    @liwenwei9
    liwenwei9
    on Sep 23, 2026 – with ChatGPT Codex Connector

    Additional reproduction on Codex Desktop for Windows:

    • Codex & OWL version: 26.917.51856 (released 2026-09-22)
    • Observed: 2026-09-23, UTC+8
    • Failing models: gpt-6-sol, gpt-6-astra, and gpt-6-luna
    • Working control: GPT-5.6
    • Reproduces in a brand-new blank task with the complete prompt: 只回复 OK ("Reply only with OK")
    • The request retried/reconnected and then failed after about 121 seconds with:
      Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt: https://platform.openai.com/docs/guides/reasoning#advice-on-prompting

    This broader reproduction suggests the issue is not limited to Astra and still occurs in the latest Windows desktop build available to this user. The harmless prompt and blank-task reproduction make prompt content and project context unlikely causes. GPT-5.6 continues to work normally in the same environment.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    appIssues related to the Codex desktop appbugSomething isn't workingmodel-behaviorIssues related to behaviors exhibited by the modelwindows-osIssues related to Codex on Windows systems

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions