False-positive cyber classification interrupts ordinary concurrency testing

What issue are you seeing?

An ordinary local software-correctness task was repeatedly blocked as
cybersecurity-related, interrupting the coding workflow and making it difficult
to continue.

The work was limited to reviewing and adding causal RED regression tests for:

• request-scoped rollback in concurrent code;
• the lifetime of an in-flight host-to-device copy during request cancellation;
• malformed serialized metadata handling.

There was no network scanning, intrusion, credential access, exploit development,
malware, persistence, external target, or security-testing request. One of the
relevant user prompts was simply:

proof RED 1 and 2

The preceding discussion concerned normal Rust/Python code review and test
coverage in an authorized local repository.

The same Codex thread recorded three failures:

• 2026-08-02 07:30:17.568 UTC
• 2026-08-02 07:32:09.484 UTC
• 2026-08-02 07:32:40.702 UTC

All three have codex_error_info: "cyber_policy" and the same message:

This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber

What steps can reproduce the bug?

1. Start a local Codex coding session in a Rust/Python repository.
2. Review a concurrency bug involving per-request ownership and an asynchronous

memory copy.

1. Ask Codex to add RED regression tests proving the two reported correctness

failures.

1. Continue the task after inspecting the existing test harness.
2. Observe three cyber_policy failures despite the request containing no

cybersecurity activity.

The likely trigger appears to be terminology such as "rollback", "corrupt
metadata", "shared block", or "memory corruption" used in an ordinary systems
programming context.

What is the expected behavior?

Codex should recognize this as normal software development and allow the selected
coding model to continue inspecting local source, editing regression tests, and
running the repository's test commands.

If an automated classifier reroutes a request, every client should show a clear
notice, provide the decision category, and make false-positive reporting
available without derailing the active task.

Additional information

• Codex: codex-cli 0.146.0
• OS: Linux 7.1.3-arch1-3 x86_64
• Surface: local Codex coding session
• Scope: authorized local repository only
• Uploaded thread: 019fb340-1b17-7a90-8e9c-0d1e3209e097
• Classification submitted through feedback/upload: safety_check
• Logs included: yes

The complete rollout log was uploaded privately through Codex's supported
feedback endpoint. It is identified by the uploaded thread ID above; it is not
attached publicly because it includes proprietary repository conversation
content.