Reachability gate (hop-by-hop)

1. Source. Application calls transporter.sendMail({ raw: <userControlled> , to: <userControlled> }) with disableFileAccess: true and/or disableUrlAccess: true configured on the transporter (forced onto mail.data in lib/mailer/mail-message.js:36-40) or per message. This is the exact scenario the flags exist for — the same precondition under which GHSA-wqvq-jvpq-h66f was accepted.
2. Guard — the access flags. For attachments the flag is enforced: a node created by _createContentNode carries disableFileAccess, so _getStream throws EFILEACCESS. Bypass: the raw branch (compile():34-35) never sets the flag on its node, so this.disableFileAccess === false and the guard at mime-node:985 / :999 is skipped. There is no other validation between mail.raw and the read; raw content shapes ({path}, {href}, stream, string, buffer) are accepted as-is by setRaw/_getStream.
3. Sink. fs.createReadStream(content.path) (file disclosure) or nmfetch(content.href, …) (SSRF). The resulting bytes are emitted as the message body by createReadStream(), which every transport pipes to its destination (smtp-transport:233, smtp-pool/pool-resource:208, ses-transport:96, sendmail-transport:184, stream-transport:67).

No guard blocks the chain; the only guard (the access flags) is structurally absent on this node.

Root cause

Inconsistent enforcement: the access policy is applied per-MimeNode via constructor options and must be re-passed at every node creation. The raw-message shortcut in compile() omits it, while all five other node builders include it. The flags are therefore enforced for every content type except the one that lets the caller supply a complete message body by path/URL.

Exploit path

Application that sandboxes untrusted mail input (disableFileAccess/disableUrlAccess set):

1. Untrusted actor supplies raw: { path: '/proc/self/environ' } (or any server file: /app/.env, key material, etc.) and to: [email protected].
2. compile() builds the raw root node without the flags; the transport reads the file and sends its contents as the message → arbitrary server-file exfiltration to an attacker-chosen mailbox.
3. Alternatively raw: { href: 'http://127.0.0.1:8080/admin' } or a cloud metadata URL → Nodemailer fetches it server-side and delivers the full response body in the email → full-response SSRF (no blind-channel limitation).

Impact

• Confidentiality (High): arbitrary local file read disclosed in the outgoing message; full-response SSRF to internal/metadata endpoints, also disclosed in the message.
• Integrity (Low): attacker-fetched/file content is injected into the delivered mail.
• The two protective flags an application relies on to contain untrusted input are silently ineffective for raw.

Preconditions

The application (a) passes disableFileAccess and/or disableUrlAccess (the documented sandboxing flags) and (b) lets untrusted input influence the raw field (and, for maximal disclosure, to). No other configuration is required; all bundled transports are affected. This mirrors the accepted precondition of GHSA-wqvq-jvpq-h66f.

Severity

• AV — message data routinely originates over the network in the apps these flags protect.
• AC — a single crafted raw object; deterministic.
• PR — the actor is a user whose input the app already treats as untrusted (the reason the flags are set); not fully anonymous in the typical deployment.
• UI — no victim interaction.
• S — impact within Nodemailer's process scope.
• C — arbitrary file read and full-response SSRF, both delivered to an attacker-chosen recipient. (The sibling jsonTransport advisory used C:L because its leak stayed in locally-returned JSON; here the bytes leave the system in the sent message, so C:H is warranted.)
• I — attacker injects fetched/file bytes into the outgoing message.
• A. Note: if a deployment fixes the recipient (to not attacker-controlled) the disclosure channel narrows and the rating degrades toward the sibling's Medium; the High rating reflects the reasonable worst case where raw and to are both untrusted.

Adversarial re-read (attempts to refute)

1. "raw content is by-design trusted, so the flags shouldn't apply." Rejected: every other content path (attachments, alternatives, html/text, icalEvent) honors the flags, and the maintainer already accepted GHSA-wqvq-jvpq-h66f for exactly this "untrusted input + flag set" model. The asymmetry — attachment {path} is blocked but raw:{path} is not — is the bug, and the PoC's CONTROL case proves the flag is otherwise effective on the same file.
2. "The raw node inherits the flags via rootNode." Rejected by code and by PoC: compile():35 constructs the node with { newline } only; MimeNode constructor sets this.disableFileAccess = !!options.disableFileAccess → false; rootNode is itself; no inheritance exists.
3. "The PoC leaks for an unrelated reason." Rejected: the CONTROL message (attachments:[{path}], same file, same transporter) returns EFILEACCESS; only the raw:{path} message leaks. The sentinel nonce exists solely in the temp file; the URL nonce is generated server-side and is only obtainable by an actual fetch. Both observables are uniquely bound to the bypass.
4. "Maybe only jsonTransport (already reported) is affected." Rejected: the PoC uses streamTransport and the root cause is in MailComposer.compile() (mailer:188), shared by all transports; jsonTransport is a different (already-fixed) path.

I could not find any guard that blocks the chain; the finding survives.

Proof of concept (safe, benign)

findings/nodemailer/raw/poc-raw-fileaccess-bypass.js — local, no network egress (loopback only), no destructive action. Output:

[CONTROL] attachment path with disableFileAccess: BLOCKED (EFILEACCESS) — flag works here
[ATTACK]  raw:{path} with disableFileAccess=true: BYPASSED — sentinel file CONTENT present in message
[ATTACK]  raw:{href} with disableUrlAccess=true (loopback server): BYPASSED — fetched body present (SSRF)
VERDICT: CONFIRMED

Run: node findings/nodemailer/raw/poc-raw-fileaccess-bypass.js (exit 0 = confirmed).

Remediation

Thread the access policy onto the raw root node, exactly as the other builders do:

if (this.mail.raw) {
    this.message = new MimeNode('message/rfc822', {
        newline: this.mail.newline,
        disableFileAccess: this.mail.disableFileAccess,
        disableUrlAccess: this.mail.disableUrlAccess
    }).setRaw(this.mail.raw);
}

(Defense in depth: setRaw/_getStream could also refuse {path}/{href} raw content when either flag is set, regardless of how the node was constructed.) Add a regression test asserting that raw:{path} and raw:{href} reject with EFILEACCESS/EURLACCESS when the flags are set, mirroring the attachment tests.