Redact All Logs

Security diagnostic output must not leak sensitive data. If you enable verbose modes (e.g., --verbose) or change log sinks (stdout vs stderr), ensure (and document) that all emitted diagnostic streams are actually redacted.

How to apply:

• Align claims with behavior: Don’t state “provider messages are redacted” unless you’ve verified it for stderr and stdout.
• Test both streams: In CI, capture both outputs and assert sensitive fields (tenant IDs, request details, raw provider messages) do not appear.
• Fail closed: If redaction can’t be guaranteed for a stream, either redact it before printing or narrow the documentation/feature to avoid misleading users.

Example (test assertion pattern):

const child = spawn('npx', ['@openai/codex-security', 'scan', '.', '--verbose']);
let stdout = '', stderr = '';
child.stdout.on('data', d => (stdout += d));
child.stderr.on('data', d => (stderr += d));

await once(child, 'close');

// Assert nothing sensitive leaked to either stream
for (const s of [stdout, stderr]) {
  expect(s).not.toMatch(/tenant/i);
  expect(s).not.toMatch(/request/i);
  expect(s).not.toMatch(/raw provider/i);
}

This prevents verbose/security diagnostics from becoming an unintended data-exfiltration channel (especially through stderr).