Nodemailer address parser keeps comments in angle-bracket addresses

## Summary Nodemailer's address parser has three published, patched fixes for RFC 5322 comment `( ... )` handling inside an addr-spec: GHSA-cc9r-2j5m-2m83 (comment glued two domain atoms into one), GHSA-g57g-f23g-4646 (the same glue for a quoted local part), and the related ReDoS fixes GHSA-prgh-xp8r-p3m5 / GHSA-v53p-9fqp-m79j. All three fixes live in `_handleAddress()`'s `joins` / `_recoverAddrSpec()` logic in `src/addressparser/index.ts`, and that logic is exercised whenever the local text-accumulation path in the tokenizer runs. That logic is **never reached at all** for the single most common way a recipient is written: a display name followed by an angle-bracket address, `Name <local@domain>`. When the input inside the angle brackets itself contains an RFC 5322 comment, e.g. ``` Legit User <[email protected](x)evil.com> ``` the comment is not recognized as an operator, is not stripped, and the parenthesis characters survive **verbatim** inside the parsed address. That value is used as-is for `envelope.to` and is written raw onto the wire in the `RCPT TO` SMTP command. I confirmed this end to end: * against nodemailer's own `jsonTransport`/`streamTransport` (`envelope.to` and the emitted `To:` header both carry the malformed value), and * against a real SMTP server (`aiosmtpd`), where the malformed value is fatal to the transaction (`501 Syntax: RCPT TO: <address> [SP <mail-parameters>]`, all recipients rejected). This reproduces unmodified on every version I tested from `6.9.16` through the current `10.0.12` HEAD, i.e. it predates and survives all three of the address-parser advisories above unchanged, because it is a different root cause in a code path those fixes never touch. ## Root cause ### 1. The tokenizer has a single-slot operator stack `src/addressparser/index.ts`, `Tokenizer.checkChar()`: ```ts } else if (!this.operatorExpecting && !this.inDomainLiteral && chr in this.operators) { this.node = { type: 'operator', value: chr }; this.list.push(this.node); this.node = null; this.operatorExpecting = this.operators[chr]; this.escaped = false; return; } ``` (currently around line 670) `operatorExpecting` is a single string, not a stack. Once `<` opens the address scope, `operatorExpecting` is set to `'>'`. The branch above that recognizes a *new* operator character (such as `(`) is gated on `!this.operatorExpecting`, which is now false. So while inside `<...>`, a `(` is never tokenized as an operator at all — it is just an ordinary character appended to the running text token, together with everything up to the next `>`. The RFC 5322 comment is not recognized, not entered, and not stripped: it rides straight through as literal text. ### 2. The `_handleAddress` state machine loses the `address` state on any operator boundary `src/addressparser/index.ts`, inside `_handleAddress()` (currently around line 365-388): ```ts switch (token.value) { case '<': state = 'address'; insideQuotes = false; break; case '(': state = 'comment'; insideQuotes = false; break; ... default: state = 'text'; insideQuotes = false; break; } ``` Even where a comment *is* recognized (outside angle brackets), closing it with `)` falls through to `default` and resets `state` to `'text'`, not back to whatever state was active before the comment opened. This is why the fixed cases (bare address, quoted local part) only work at all: the leaked comment tail lands in the free-`text` bucket, where `_recoverAddrSpec()`/the ADDR_SPEC extraction logic can find and strip it. Inside `<...>`, none of that applies in the first place (point 1), so there is no text-bucket safety net to fall back on — the whole point 1/point 2 interaction is what makes the angle-bracket case categorically different from, and unprotected by, the three existing fixes. ### 3. The one existing safety net requires whitespace, and this value has none `_recoverAddrSpec()` (currently around line 283) is the function all three prior fixes rely on to peel a clean addr-spec back out of a polluted value. Its very first line is: ```ts function _recoverAddrSpec(data: { address: string; text: string }): void { if (!HAS_WHITESPACE.test(data.address)) { return; } ... ``` `[email protected](x)evil.com` contains no whitespace at all — the comment content `x` had none, and nothing in the angle-bracket path inserts any — so `_recoverAddrSpec` bails out immediately and never even looks at the value. There is no cleanup path for this shape. ## Proof of Concept Environment: Node.js, current `master` (commit `30e4cfd5902fd5da4d1bc76dc8e4bc92d288adef`, package.json version `10.0.12`), no special transport configuration required. ### Parser-level confirmation ```js const addressparser = require('nodemailer/lib/addressparser'); console.log(addressparser('Legit User <[email protected](x)evil.com>')); // [ { address: '[email protected](x)evil.com', name: 'Legit User' } ] ``` Compare with the already-fixed, unbracketed form of the exact same bytes: ```js console.log(addressparser('[email protected](x)evil.com')); // [ { address: '[email protected]', name: 'evil.com' } ] <- correctly stripped ``` ### `sendMail` / envelope confirmation (jsonTransport, no network) ```js const nodemailer = require('nodemailer'); (async () => { const t = nodemailer.createTransport({ jsonTransport: true }); const info = await t.sendMail({ from: '[email protected]', to: 'Legit User <[email protected](x)evil.com>', subject: 'hi', text: 'x' }); console.log('envelope.to:', info.envelope.to); })(); ``` Observed output: ``` envelope.to: [ '[email protected](x)evil.com' ] ``` The raw `streamTransport` message shows the same value in the emitted header: ``` From: [email protected] To: Legit User <[email protected](x)evil.com> Subject: hi ``` ### Wire-level confirmation against a real SMTP server ```js const net = require('net'); const nodemailer = require('nodemailer'); const server = net.createServer(sock => { let buf = ''; sock.write('220 sink ESMTP\r\n'); sock.on('data', d => { buf += d; let i; while ((i = buf.indexOf('\r\n')) >= 0) { const line = buf.slice(0, i); buf = buf.slice(i + 2); const u = line.toUpperCase(); if (u.startsWith('EHLO')) sock.write('250-sink\r\n250 8BITMIME\r\n'); else if (u.startsWith('MAIL')) sock.write('250 ok\r\n'); else if (u.startsWith('RCPT')) { console.log('wire RCPT TO :', line); sock.write('250 ok\r\n'); } else if (u.startsWith('DATA')) sock.write('354 go\r\n'); else if (line === '.') sock.write('250 ok\r\n'); else if (u.startsWith('QUIT')) { sock.write('221 bye\r\n'); sock.end(); } else sock.write('250 ok\r\n'); } }); }); server.listen(0, '127.0.0.1', async () => { const t = nodemailer.createTransport({ host: '127.0.0.1', port: server.address().port, secure: false }); const info = await t.sendMail({ from: '[email protected]', to: 'Legit User <[email protected](x)evil.com>', subject: 'hi', text: 'x' }); console.log('envelope.to:', info.envelope.to); t.close(); server.close(); }); ``` Observed: ``` wire RCPT TO : RCPT TO:<[email protected](x)evil.com> envelope.to: [ '[email protected](x)evil.com' ] ``` The same recipient string sent at a real, conformant SMTP server (`aiosmtpd`, a widely used reference-grade Python SMTP implementation) instead of the sink above: ``` Error: Can't send mail - all recipients were rejected: 501 Syntax: RCPT TO: <address> [SP <mail-parameters>] code: 'EENVELOPE', responseCode: 501, command: 'RCPT TO', rejected: [ '[email protected](x)evil.com' ] ``` So against a strict receiver the immediate, concretely demonstrated outcome is that the message is bounced: an ordinary-looking recipient address that a sender/application believes is valid (and, see below, that some app-side validators *do* read as valid, at `good-corp.com`) silently fails to send. ### Interpretation-conflict confirmation (same class as GHSA-cc9r-2j5m-2m83) The same string, read by two other parsers an application might use to gate on the recipient's domain, produces two more, mutually different readings: ```python >>> import email.utils >>> email.utils.parseaddr('Legit User <[email protected](x)evil.com>') ('Legit User (x)', '[email protected]') ``` ```python >>> import email.policy, email.parser >>> msg = email.parser.Parser(policy=email.policy.default).parsestr( ... 'To: Legit User <[email protected](x)evil.com>\n\n') >>> addr = msg['To'].addresses[0] >>> addr.username, addr.domain ('user', 'good-corp.com') ``` So for the exact same bytes: | Parser | Domain read | |---|---| | Python `email.utils.parseaddr` | `good-corp.comevil.com` (attacker-registrable) | | Python `email.policy.default` (defects not inspected) | `good-corp.com` (clean, would pass an allow-list check) | | Nodemailer `addressparser` (this report) | `good-corp.com(x)evil.com` (illegal characters, rejected by a strict MTA) | This is the exact "which parser sees what" comparison table from GHSA-cc9r-2j5m-2m83, reproduced through a different, unpatched code path: an application validating with `email.policy.default` without checking `.defects` (a documented pattern in that advisory) approves `good-corp.com` and hands the string to Nodemailer, which then emits a different, malformed value on the wire. ### Version range The built `addressparser`/`sendMail` output was tested for this exact input across published releases: | Version | `envelope.to` | |---|---| | `6.9.16` | `[email protected](x)evil.com` | | `9.0.6` | `[email protected](x)evil.com` | | `9.1.0` (fixed GHSA-cc9r-2j5m-2m83) | `[email protected](x)evil.com` | | `10.0.1` | `[email protected](x)evil.com` | | `10.0.12` / current HEAD (fixed GHSA-g57g-f23g-4646) | `[email protected](x)evil.com` | Unchanged across every one of the three address-parser advisory fixes, because none of them touch the tokenizer's single-slot `operatorExpecting` state or the `_handleAddress` state-machine reset on `)`. ## Impact * **Confirmed, demonstrated impact:** a recipient address that contains a perfectly legal RFC 5322 comment inside angle brackets — the single most common way a `To`/`Cc`/`Bcc`/`From`/`Reply-To` value is written (`Display Name <user@domain>`) — is parsed into a value containing SMTP-illegal characters (`(`, `)`) that a conformant receiving SMTP server rejects outright, silently failing message delivery to what the sender/application believed was a valid, already-validated recipient (CWE-20, availability/reliability impact on outbound mail). * **Demonstrated interpretation conflict (CWE-436):** the malformed value that reaches `envelope.to` and the wire `RCPT TO` command disagrees with how other parsers, including ones explicitly called out as an at-risk validation pattern in GHSA-cc9r-2j5m-2m83 (Python's `email.policy.default` without defect inspection), read the exact same bytes. An application gating outbound mail on "is the recipient's domain in my allow-list" using such a validator can approve a value that Nodemailer then sends corrupted, and depending on the receiving MTA's own comment handling, a different domain than either the sender or the validator believed was being used. * **Reachability:** every code path that ends up calling `MimeNode._parseAddresses()` — `to`, `cc`, `bcc`, `from`, `replyTo`, and the explicit `envelope` option — is affected identically, since they all funnel through the same `addressparser()`/`_handleAddress()` logic. ## Suggested fix The underlying issue is architectural: `Tokenizer.operatorExpecting` is a single value, so it cannot represent "currently inside an angle-address, and also currently inside a comment nested within it." A minimal fix is to make the tokenizer maintain a stack of expected closing characters (pushing `')'` when entering a comment while another operator scope, e.g. `'>'`, is already open, and popping back to the outer expectation on `)`), so a comment can be recognized and stripped no matter what scope it is nested inside. `_handleAddress`'s state machine would then need to restore the previous `state` on a comment close instead of unconditionally resetting to `'text'`, so recognized-but-stripped comment content inside `<...>` continues to accumulate as `address` rather than leaking into `text`. As defense in depth, `_recoverAddrSpec()` (or an equivalent check run specifically on `data.address` for the angle-bracket case) should also reject/clean any parsed address value that still contains RFC 5322 special characters such as `(`/`)` regardless of whether whitespace is present, since `envelope.to` is used directly to build an SMTP command and any character outside `Mailbox` (RFC 5321 §4.1.2) reaching that string is already the wrong outcome, independent of how it got there. ## Regression test suggestion `Name <[email protected](x)evil.com>` and the bcc/cc/from/replyTo/envelope equivalents should all parse to `{ address: '[email protected]', name: 'Name evil.com' }` or be rejected outright by `sendMail`, never silently carry `(`/`)` into `envelope.to`.