Environment
@openai/codex-security 0.1.10 (npm), bundled plugin 0.1.17- Invocation:
codex-security scan <repo> --mode standard --headless --auth api-key --output-dir <dir> --scan-prompt-file <file> --format json (also reproduced with --model gpt-5.6-terra --effort high)
- Auth:
OPENAI_API_KEY (organization API key), unattended CI usage
- Runtime: Linux container (read-only rootfs, seccomp allowlist per the published sandbox requirements), Apple Silicon host
- Scan target: a private, first-party enterprise codebase (an Electron IM desktop client, TypeScript/Vue/React, ~3,700 files). We own the code; the scan is part of our merge-request security review pipeline.
What happened
The scan enumerated 3,743 files and ran normally for ~13 minutes. After ~31M input tokens
(estimated cost $26.7) and 11 files reviewed, the scan aborted:
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
Partial output was kept at /state/out.
- Exit code:
2
- stdout: empty / not JSON, despite
--format json
- The moderation notice goes to stderr as plain text only
Why this is a problem
- A security scanner being blocked by a cybersecurity-risk filter is self-defeating.
The tool's entire purpose is authorized security review of our own code. The repo
contains ordinary product code (including standard attachment-encryption/decryption
logic, which we suspect triggered the classifier).
- No structured failure envelope. With
--format json we expect every terminal
state — including a moderation abort — to produce a JSON envelope with a status and
diagnostic code. Instead automation sees exit 2 + non-JSON stdout, indistinguishable
from a crash. Our pipeline surfaced it as invalid_json with no cause attached.
- The abort happened 13 minutes and $26 in. If the request/prompt was going to be
flagged, evaluating that before burning ~31M input tokens would save real money; at
minimum, the partial results that were "kept" should be sealed into the JSON output.
Questions / requests
- Please emit a structured JSON error envelope (status + diagnostic code, e.g.
moderation_blocked) on moderation aborts when --format json is set.
- Please evaluate the flag earlier in the scan lifecycle, or refund/seal partial
results when a scan is terminated by moderation after significant spend.
- Guidance on what content classes trigger the flag would help us pre-scope scans
(e.g. exclude crypto helper directories) .
Environment
@openai/codex-security0.1.10 (npm), bundled plugin 0.1.17codex-security scan <repo> --mode standard --headless --auth api-key --output-dir <dir> --scan-prompt-file <file> --format json(also reproduced with--model gpt-5.6-terra --effort high)OPENAI_API_KEY(organization API key), unattended CI usageWhat happened
The scan enumerated 3,743 files and ran normally for ~13 minutes. After ~31M input tokens
(estimated cost $26.7) and 11 files reviewed, the scan aborted:
2--format jsonWhy this is a problem
The tool's entire purpose is authorized security review of our own code. The repo
contains ordinary product code (including standard attachment-encryption/decryption
logic, which we suspect triggered the classifier).
--format jsonwe expect every terminalstate — including a moderation abort — to produce a JSON envelope with a status and
diagnostic code. Instead automation sees exit 2 + non-JSON stdout, indistinguishable
from a crash. Our pipeline surfaced it as
invalid_jsonwith no cause attached.flagged, evaluating that before burning ~31M input tokens would save real money; at
minimum, the partial results that were "kept" should be sealed into the JSON output.
Questions / requests
moderation_blocked) on moderation aborts when--format jsonis set.results when a scan is terminated by moderation after significant spend.
(e.g. exclude crypto helper directories) .