codex - 💡(How to fix) Fix False-positive safety block when migrating legitimate biomedical research workflow skill files

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

Latest

What subscription do you have?

Free

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

I was evaluating a migration from Antigravity to Codex. This happened while ChatGPT 5.5/Codex was migrating my existing workflows into Codex skill files.

I do serious biomedical research involving vaccine design, drug discovery, wet-lab prediction, and Alzheimer’s-related compound discovery. These are expert-supervised research workflows, not consumer medical advice. The task was not “tell a patient what to take” or “give me clinical treatment guidance.” It was migration and organization of existing research workflows into Codex-compatible skill files.

Instead, the system blocked the task with:

Invalid prompt: we've limited access to this content for safety reasons. This type of information may be used to benefit or to harm people. We are continuously refining our work in this area, and you can read more about our approach in our blog post...

This does not appear to be protecting anyone. It appears to be a blanket keyword/content-based restriction over long biomedical workflow files. Serious biomedical workflows naturally contain terms involving compounds, assays, targets, disease biology, wet-lab validation, toxicity, organoids, and protocol planning. Blocking those terms during file migration makes Codex unusable for precisely the kind of high-value scientific work it should support.

The practical effect is that legitimate research groups cannot use Codex for biomedical workflow management, even when the work is expert-led, preclinical, computational, and infrastructure-oriented. Meanwhile, researchers can use Antigravity, Cursor, Windsurf, open models, local models, or unrestricted systems for similar work. Blanket blocking does not meaningfully improve safety; it mainly prevents serious labs from using OpenAI’s tools.

Expected behavior:

• Codex should distinguish between user-provided file/workflow migration and generation of unsafe biological instructions.
• Codex should allow organization, refactoring, summarization, and conversion of existing biomedical research workflows into skill files.
• Codex should distinguish consumer clinical advice from expert-supervised preclinical research workflow support.
• If a specific section is blocked, Codex should identify the triggering section and explain the precise policy issue instead of failing opaquely.
• OpenAI should consider a more contextual research mode for legitimate biomedical users rather than blanket syntax-based restrictions.

Actual behavior: Codex blocked the workflow migration simply because the files contained biomedical research content.

Impact: This makes Codex impractical for serious biomedical research teams. It prevents use cases that could include computational prediction, compound prioritization, Alzheimer’s research, wet-lab planning, organoid validation planning, reproducible analysis, and research workflow governance.

Suggested improvement: Replace blanket keyword-based blocking with context-aware safeguards. For example:

• allow user-provided biomedical workflow files to be migrated, organized, and converted into skills;
• gate genuinely dangerous requests based on intent and operational specificity, not isolated terms;
• provide section-level explanations when something is blocked;
• support expert-supervised preclinical research workflows with appropriate warnings, logging, and review expectations rather than refusing the whole task.

As implemented, this feels like security theater at the cost of making OpenAI models unusable for potentially high-impact biomedical research.

Just a little perspective.

What steps can reproduce the bug?

1. Open Codex / ChatGPT 5.5 in the Codex App.
2. Ask it to migrate existing Antigravity workflows into Codex skill files.
3. Provide a long user-owned biomedical research workflow that contains terms related to computational prediction, drug discovery, vaccine design, Alzheimer’s research, wet-lab validation planning, organoids, targets, assays, toxicity, and compound prioritization.
4. Make clear that the task is file/workflow migration, not clinical advice or protocol generation.
5. The system blocks the request with the “Invalid prompt” safety message.

Minimal representative example:

Please convert the following existing internal workflow into a Codex skill file.

Scope:
- computational prediction
- biomedical literature analysis
- compound prioritization
- Alzheimer’s research
- drug discovery
- vaccine design
- wet-lab validation planning
- organoid model planning

Safety boundary:
- do not provide patient-specific medical advice
- do not recommend treatment to a patient
- do not generate new operational biological instructions
- only migrate and organize this user-provided workflow into a skill file

Task:
Create the Codex SKILL.md structure for this workflow.

### What is the expected behavior?


**What is the expected behavior?**
```markdown
Codex should allow migration, organization, and refactoring of user-provided biomedical research workflow files into Codex skill files.

It should distinguish between:
- file migration / workflow conversion
- computational research support
- expert-supervised preclinical research planning
- consumer clinical advice
- generation of unsafe operational biological instructions

If a specific part of the workflow is not allowed, Codex should identify the specific section and explain the issue. It should not reject the entire task based on broad biomedical terminology.

### Additional information

This is especially damaging because Codex skills are meant to capture complex workflows. Long biomedical workflows naturally contain sensitive domain vocabulary, even when the task is infrastructure, documentation, or computational research support.

The current behavior makes Codex hard to evaluate for serious research groups. It also creates an incentive to use less restricted tools or local/open models instead of OpenAI products.

A better product behavior would be a contextual research mode or file-migration mode that permits user-provided biomedical workflows to be copied, organized, summarized, modularized, and converted into skills while still preventing unsafe requests for patient treatment instructions or harmful biological execution details.

I am reporting this because I was actively considering moving workflows from Antigravity to Codex, and this block happened during that migration evaluation.