False-positive cyber_policy repeatedly terminates benign PS5 GPU driver/compiler work

What version of Codex CLI is running?

codex-cli 0.145.0

What subscription do you have?

ChatGPT Pro

Which model were you using?

gpt-5.6-sol with high reasoning effort

What platform is your computer?

Darwin 25.5.0 arm64 arm

What terminal emulator and version are you using (if applicable)?

Apple Terminal 470.2

Codex doctor report

Redacted public summary:

{
  "overallStatus": "warning",
  "codexVersion": "0.145.0",
  "relevantChecks": {
    "auth.credentials": "ok",
    "config.load": "ok",
    "installation": "ok",
    "network.provider_reachability": "ok",
    "updates.status": "ok"
  },
  "unrelatedWarnings": [
    "background app-server socket stale or unreachable",
    "Responses WebSocket authentication warning; HTTPS remained reachable"
  ]
}

The full doctor report is omitted from this public issue because it contains local filesystem and authentication-configuration details. It can be provided privately if needed.

What issue are you seeing?

A long-running Codex thread performing benign, local PS5 graphics-driver and shader-compiler development was repeatedly terminated by the cybersecurity classifier.

Uploaded thread: 019f98fe-d6d5-7b52-b55f-b4cec1c1f158

The work is an open-source clean-room implementation of a PS5 graphics API and shader compiler. It involves C code, Mesa/AMD GPU source references, command-buffer register programming, host tests, and validation on hardware owned by the user. It does not involve network intrusion, credential theft, malware, scanning, persistence, exfiltration, or unauthorized access.

The thread contains 19 task_complete failures with codex_error_info: "cyber_policy" between 2026-07-25T18:46:55Z and 2026-07-26T02:50:09Z. The failed turns accumulated approximately 6,793,870 ms (113 minutes) before being terminated.

Representative error:

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

Representative affected turn IDs:

• 6367df66-5541-4ca0-9fe2-b3d5e2d6e00a
• aa2b40d7-1c22-42e3-99cd-243d5668125b
• 1a94488d-2bf3-46db-8acf-9158c345f84c
• e9dc638e-9fd0-4dab-b56a-0153435b652a

The failures happened after ordinary source inspection, compiler work, test execution, and hardware graphics validation. Repeated automatic goal-continuation turns encountered the same classifier failure, interrupting progress and losing substantial computation time.

What steps can reproduce the bug?

1. Open the uploaded thread 019f98fe-d6d5-7b52-b55f-b4cec1c1f158.
2. Work in a local C project implementing a clean-room PS5 graphics library and Mesa/ACO-based shader compiler.
3. Inspect local Mesa/Linux AMD driver sources and emulator references to compare gfx1013 NGG geometry-shader register state.
4. Build host/prospero targets and deploy a graphics sample to a user-owned PS5 over the local LAN.
5. Continue debugging an NGG geometry shader that submits successfully but produces no rasterized pixels.
6. Observe turns terminating with codex_error_info: "cyber_policy", despite the absence of offensive-security activity.
7. Resume the same benign goal and observe the classification recurring across later turns.

What is the expected behavior?

Low-level graphics-driver, firmware-compatibility, emulator, and shader-compiler development on user-owned hardware should be classified according to its actual intent rather than being treated as high-risk cyber activity because it contains words such as firmware, reverse engineering, credentials, memory, registers, or bypass in legitimate platform code.

If a safety review is required, it should happen before a turn consumes several minutes of computation. The client should preserve completed progress, identify the specific request that triggered review, and avoid repeatedly terminating benign continuation turns in the same thread.

Additional information

Related issue #22554 reports false positives during defensive security review. This report is distinct: the affected workflow is ordinary GPU/compiler/platform engineering rather than penetration testing or a Trusted Access security workflow.

The uploaded thread contains the full prompts, tool calls, task-completion errors, and local engineering context needed to investigate the classifier decisions without posting proprietary or sensitive files publicly.