Codex Security estimated $5.72 for a four-line diff, exceeded its $5 limit, then failed

[stephenstarkie]

I evaluated @openai/codex-security version 0.1.4 in GitLab CI against a committed four-line YAML diff: three inserted lines and one deleted line.

The security scan ran for 2 minutes 43 seconds. Its reported estimated cost rose from $0.10552 after nine seconds to $5.7191285, exceeding the configured $5.00 maximum by approximately 14%. It then stopped without producing a sealed scan manifest, and the CI job failed.

[02:39] Estimated cost: $4.5106495 of $5.00 limit
[02:43] Estimated cost: $5.7191285 of $5.00 limit
codex-security: Scan stopped: estimated cost $5.7191285 exceeded the $5.00 limit
Codex Security did not produce a sealed scan manifest.

The ordinary Codex code review of the same diff completed in approximately 32 seconds. I removed the containerised review job immediately.

Why this matters:

• The input was a four-line infrastructure change.
• The configured spending limit did not behave as a hard ceiling.
• The estimate increased from $3.334205 to $5.7191285 in six seconds.
• The command failed without a usable sealed result.
• Routine per-commit execution could create substantial, unpredictable API expenditure.

Requested improvements:

1. Enforce a genuine hard spending ceiling.
2. Provide a meaningful cost estimate before paid work begins.
3. Document realistic per-scan cost ranges.
4. Provide a low-cost mode suitable for routine CI diffs.
5. Produce a usable partial report when a limit is reached.

Full redacted timeline and reproduction details: https://gitlab.com/stephenstarkie/codex-security-cost-incident

All dollar figures are estimates reported by codex-security itself, not asserted final billed charges.