openai / openai/codex

Repeated cyber-safety false positives during benign scientific-software audit, persisting after prompt rewording

Open
#40,421 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

bug CLI safety-check
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

What version of Codex CLI is running?

0.149.0

What subscription do you have?

ChatGPT Plus

Which model were you using?

gpt-5.6-sol, Ultra reasoning

What platform is your computer?

macOS

What issue are you seeing?

Codex repeatedly displays:

This content can't be shown
We take extra caution with cybersecurity requests... Trusted Access

during a completely benign, local, read-only scientific-software audit.

The repository implements a sports-trajectory coordinate system. The audit is checking whether missing or unverified vertical (z) metadata can incorrectly acquire governed 3D scientific authority.

There is no network target, penetration testing, credential access, malware, exploitation, persistence, or third-party system involved.

The warning has now occurred five times in the same audit workflow.

Initially, the audit terminology included words such as red team, authorization, forged, tamper, and adversarial, used strictly in a scientific/software-integrity sense.

I then deliberately reworded subsequent prompts to avoid obvious cybersecurity terminology, but the cyber-safety warning continued to trigger.

Codex itself appears to understand the scientific context and continues the audit correctly after the blocked narration. This suggests the false positive may occur in a separate safety/display layer, potentially using accumulated session or generated-output context rather than only the current user prompt.

What steps can reproduce the bug?
  1. Run a local read-only audit of scientific software that validates 3D trajectory metadata.
  2. Test whether caller-provided or insufficiently verified z metadata can become scientifically authoritative downstream.
  3. Use bounded synthetic test cases only; no network or external systems are involved.
  4. During the audit, Codex generates software-integrity terminology around validation, authority, altered metadata, and adversarial test cases.
  5. Observe the cybersecurity Trusted Access warning.
  6. Reword subsequent prompts to explicitly use scientific-software terminology and avoid cybersecurity terminology.
  7. Observe that the warning continues to recur.
What is the expected behavior?

A benign scientific-software audit should be classified according to its actual domain and requested capabilities rather than isolated security-adjacent vocabulary.

Rephrasing a prompt to remove such terminology should also prevent repeated false positives if the current request itself is benign.

Actual behavior

The cybersecurity warning has appeared five times and hides portions of Codex's output despite the underlying task remaining a local scientific validation audit.

Additional information

This appears related to #34306, #32597, #34913, and #36671, but differs in that the affected task is not cybersecurity work at all.

In practical terms, the audit is asking:

“Can an unverified z-coordinate incorrectly become scientifically authorized as 3D data?”

and the safety layer repeatedly interprets the surrounding validation terminology as cybersecurity.

I have redacted screenshots available that show the repeated Trusted Access warnings and the surrounding benign scientific-software context.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

No source file, test, or entry point is named. Start by reproducing the behavior with Codex CLI 0.149.0 on macOS using the bounded local scientific audit, then compare the initial and reworded prompts. Done means benign scientific-audit output is not repeatedly hidden by the cybersecurity warning.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust
Domain
ai, cli
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.