Windows scan output directories are not ACL-hardened like credential homes

[batmnnn]

Summary

On POSIX, scan output directories are required to be mode 0700 and owned by the current user, and shared parents are checked for sticky/trusted ancestry (#118). Credential homes on Windows get an equivalent current-user-only ACL via secureWindowsCredentialHome.

Scan output on Windows has no equivalent. requirePrivateOutputDirectory and requireSecureOutputAncestry both early-return on win32, so prepared scan trees keep inherited ACLs. On a shared Windows directory (e.g. C:\Users\Public\... or a folder with inherited Allow rules for other users), another local principal can read or replace scan artifacts between prepare and complete-scan / contract load.

Evidence (current main @ 3bf2621, @openai/codex-security@0.1.5)

sdk/typescript/src/runtime.ts:

• requirePrivateOutputDirectory — if (process.platform === "win32") return;
• requireSecureOutputAncestry — same early return
• secureWindowsCredentialHome / private ACL apply+verify exists for credential homes only
• validatePreparedOutputDir / validateOutputDir / contract requireScanRoot call the no-op Windows helpers
• Multiscan ensureOutputDirectory previously only mkdir'd without any private check

Workbench Python require_canonical_scan_directory also skips privacy on nt.

Steps to reproduce (Windows)

1. Create a scan output path under a directory that inherits Allow ACEs for other local users (e.g. a Public or shared folder).
2. Run codex-security scan (or bulk-scan) with that --output-dir.
3. Observe the created scan directory does not have access-rule protection restricted to the current user (contrast with …/codex-home after login, which does).
4. Another local user who inherits Allow on the parent can read findings/reports or replace sealed artifacts before completion.

Why in scope (SECURITY.md)

• Credentials, private source, and scan results must stay out of other security principals
• File-replacement races that forge completed scans or leak results
• Same boundary already enforced for Windows credential homes and POSIX scan output

Suggested fix

Reuse the credential-home Windows ACL helper for scan output:

1. Apply + verify a current-user-only ACL when preparing scan output
2. Re-verify on contract load / scan-root checks
3. Apply the same private-output gate to multiscan / bulk-scan campaign roots

Affected version

• @openai/codex-security@0.1.5
• Confirmed on main at 3bf2621