tornado: CurlAsyncHTTPClient enforces no response-size limit — decompression bomb drives unbounded memory accumulation to OOMGHSA-chx6-46f5-w4vp

An unbounded memory accumulation (decompression bomb) in tornado.curl_httpclient.CurlAsyncHTTPClient — the client-side sibling gap of CVE-2026-49855 — verified end-to-end on the 2026-08-15 master snapshot (6.6.dev1) and present unchanged in the latest release tag v6.5.8 and on master (checked 2026-08-17). When a Tornado application configures the curl client (the documented deployment for proxy support / advanced TLS options) and fetch()es an attacker-chosen or attacker-compromised URL with default decompress_response=True, a malicious server replying Content-Encoding: gzip with a ~2.8 MB wire bomb drove the client's RSS from 30,884 kB to 1,032,100 kB (~1008 MB) in 3.18 s (~350 MB/s, monotonic, no plateau) until the kernel OOM-killed the process (exit 137, cgroup OOMKilled=true) — with the transfer only 67% complete and no client-side size check ever intervening: curl_httpclient.py contains zero occurrences of max_body_size/MAXFILESIZE. The identical bomb against the default SimpleAsyncHTTPClient fails cleanly at ~65 MB, because every size gate CVE-2026-49855 added (compressed CL, chunked total, cumulative decompressed size — http1connection.py:620,676,742) lives in code the curl client never executes. This is a distinct component from the published advisory (which fixed _GzipMessageDelegate/SimpleAsyncHTTPClient only) and from the other curl-client advisories (credential handle-reuse GHSA-pw6j-qg29-8w7f, header CRLF GHSA-w235-7p84-xx57); the file's full commit history (latest 2026-06-17) shows no response-size work.

Details

tornado/curl_httpclient.py (line numbers identical on master 6.6.dev1, v6.5.8, and the audited snapshot):

"buffer": BytesIO(),                                  # :202 — plain BytesIO, no accounting
...
else:
    write_function = buffer.write                     # :359 — every decompressed byte lands here
curl.setopt(pycurl.WRITEFUNCTION, write_function)     # :360
...
if request.decompress_response:                       # default True (HTTPRequest)
    curl.setopt(pycurl.ENCODING, "gzip,deflate")      # :373-374 — libcurl advertises + auto-decodes

libcurl decompresses the response before invoking WRITEFUNCTION, so the callback receives decompressed bytes, which are appended to an unbounded BytesIO until the transfer ends or the process dies. The only ceiling is request_timeout (default 20 s) — at zlib's hundreds of MB/s that still permits many GB of accumulation; the PoC raised it to 300 s and the 1 GiB cgroup cap was hit in 3.2 s regardless. There is no pycurl.MAXFILESIZE, no max_buffer_size/max_body_size plumbing (the constructor accepts no body-size option), and streaming_callback users fare no better (the callback variant at :353-356 also performs zero accounting).

Contrast — SimpleAsyncHTTPClient path (tornado/http1connection.py), all absent from the curl path:

if cast(int, content_length) > self._max_body_size:        # :620  Content-Length gate
if total_size > self._max_body_size:                       # :676  chunked total gate
if self._decompressed_body_size > self._max_body_size:     # :742  CVE-2026-49855 decompressed gate

SimpleAsyncHTTPClient.initialize() defaults max_buffer_size = 104857600 (100 MiB) with max_body_size defaulting to it (simple_httpclient.py:117-121); CurlAsyncHTTPClient.initialize() has no corresponding parameter at all.

Attack chain (attacker = malicious HTTP server; victim = any Tornado app doing fetch() on attacker-influenced URLs — URL fetchers, webhook processors, link previewers, RSS/probe pollers):

1. App configures AsyncHTTPClient.configure("tornado.curl_httpclient.CurlAsyncHTTPClient") (documented for proxy support; proxies are only supported with the curl client).
2. App calls fetch("http://attacker/...") with defaults → request advertises Accept-Encoding: gzip,deflate.
3. Attacker replies 200, Content-Encoding: gzip, Transfer-Encoding: chunked, body = a gzip stream of zeros (4,174,525 wire bytes expanding to 4 GiB, 1029:1, sent in 64 KB chunks).
4. libcurl auto-decodes at ~350 MB/s into buffer.write with no size accounting → process RSS climbs linearly until OOM. The response need not complete: the client died with 2,818,048/4,174,525 wire bytes delivered (67%).

Variant without compression: decompress_response=False plus an endless streaming body (no Content-Length, no final chunk) feeds the same unaccounted buffer.write — this client never enforces any cap on any path.

PoC

Verified end-to-end 2026-08-15 in a single container (cgroup --memory 1g --memory-swap 1g so the exhaustion endpoint is safe and fast to observe), two processes over a real TCP socket on 127.0.0.1:8081: a raw-socket malicious server (evil_server.py, builds the 4 GiB-of-zeros gzip bomb once, serves it with chunked framing) and the Tornado victim (victim_curl.py, configures CurlAsyncHTTPClient, fetch(..., request_timeout=300), samples /proc/self/status VmRSS every 0.2 s from a monitor thread so the curve survives the OOM kill).

Build & run the victim

git clone https://github.com/tornadoweb/tornado
cd tornado
pip install pycurl
python evil_server.py &     # 127.0.0.1:8081; prints BOMB_BUILT wire_bytes=... ratio=... EVIL_LISTENING
python victim_curl.py       # victim: CurlAsyncHTTPClient fetch -> expect linear RSS rise -> OOM
python victim_simple.py     # control: default SimpleAsyncHTTPClient on the same bomb

Verified environment: debian:bookworm-slim, Python 3.11.2, python3-pycurl 7.45.2 (libcurl 7.88.1, zlib 1.2.13), tornado master snapshot 6.6.dev1 of 2026-08-15 run from the source tree (sys.path), container memory capped at 1 GiB. curl_httpclient.py verified byte-equivalent (still zero size-limit references) in tag v6.5.8 and on master as of 2026-08-17.

Reproduction steps

1. Precondition — the documented curl-client deployment fetching a remote URL. The vulnerability requires the application to use CurlAsyncHTTPClient (the standard configuration when proxy support or advanced TLS options are needed) with default decompress_response=True, and to fetch a URL whose server the attacker controls or has compromised. victim_curl.py implements exactly that (AsyncHTTPClient.configure("tornado.curl_httpclient.CurlAsyncHTTPClient") → fetch()), and the server log confirms the ENCODING path engaged — the victim's request arrived with User-Agent: Mozilla/5.0 (compatible; pycurl) and Accept-Encoding: gzip,deflate.

2. Attack: start evil_server.py, wait for EVIL_LISTENING, then run victim_curl.py (the fetch itself is the attack; no further interaction).

3. Expected: victim_curl_rss.log shows VmRSS 30,884 → 1,032,100 kB over 3.18 s, rising ~350 MB/s with no plateau; the container kills the victim (exit 137, docker OOMKilled=true); the server logs CLIENT_DIED_MID_TRANSFER wire_sent=2818048/4174525 — the accumulation is bounded only by available memory, never by a client-side check, and the 4 GiB payload was never fully delivered.

4. Variant: with decompress_response=False and an unterminated chunked body (server keeps sending forever), the same buffer.write path accumulates unbounded plain bytes — no compression needed; request_timeout only extends the ceiling.

5. Control: victim_simple.py fetches the identical bomb with the default SimpleAsyncHTTPClient → clean FETCH_FAILED: HTTP 599: Connection closed, RSS peak ~65 MB (24,984 → 66,712 kB), script exit 0 — the CVE-2026-49855 accounting in http1connection.py aborts the transfer, proving the gap is specific to the curl client.

PoC source

Full PoC source (victim_curl.py, stdlib only, no dependencies): victim_curl.py (secret gist, unlisted).

The gist carries evil_server.py (bomb server), victim_curl.py (victim), victim_simple.py (control), and REPRODUCE.md.

Captured output (2026-08-15 run, verbatim excerpts)

evil_server.log:
BOMB_BUILT wire_bytes=4174525 decompressed_bytes=4294967296 ratio=1029:1
EVIL_LISTENING 127.0.0.1:8081
REQUEST_FROM 127.0.0.1:57544 -> GET /bomb HTTP/1.1
REQUEST_HEADERS:
GET /bomb HTTP/1.1
Host: 127.0.0.1:8081
User-Agent: Mozilla/5.0 (compatible; pycurl)
Accept: */*
Accept-Encoding: gzip,deflate
WIRE_SENT 65536/4174525
WIRE_SENT 2162688/4174525
CLIENT_DIED_MID_TRANSFER wire_sent=2818048/4174525 err=ConnectionResetError(104, 'Connection reset by peer')

victim_curl_rss.log (VmRSS kB, every 0.2 s):
0.00  30884
0.61  231848
1.41  514720
2.22  794480
2.82 1000756
3.18 1032100        <- last sample before kill

driver_f1.log:
timeout 240 python3 /e2e/F1/victim_curl.py ...  606 Killed
VICTIM_CURL_EXIT=137         # docker inspect -> OomKilled: true

victim_simple.log (control, same bomb):
FETCH_FAILED: HTTP 599: Connection closed
SCRIPT_FINISHED_NORMALLY     # exit 0, VmRSS peak 66712 kB

Honest framing of the endpoint: the OOM kill at ~1008 MB was forced by the test's 1 GiB cgroup cap as the observation instrument; the "unbounded" claim rests on the linear no-plateau RSS curve, death at 67% wire delivery, and the absence of any size accounting in the code path — with more memory the transfer would have continued to the full 4 GiB payload.

Impact

Denial of service (memory exhaustion) of any Tornado application that uses the curl HTTP client and fetches attacker-influenced URLs. A single ~4 MB response kills a 1 GiB process in ~3 s; wire cost scales as available_memory / 1000. Because the accumulation happens on the shared event loop's client, one malicious response takes down the entire application (all concurrently served users), and a slow endless-body variant drains memory gradually below detection thresholds. The attack requires no privileges, no user interaction, and only that the victim's configured client visits the attacker's origin.

Suggested fix

Enforce a byte budget in the write path of CurlAsyncHTTPClient: wrap the WRITEFUNCTION (both the buffer.write branch and the streaming_callback branch) in a counter that aborts the transfer (curl.setopt(pycurl.FAILONERROR)-style cancellation or raising from the callback) once the received total exceeds max_body_size, plumbed from initialize() with the same 100 MiB default as SimpleAsyncHTTPClient. Because libcurl decompresses before the write callback, the counter naturally measures decompressed bytes — the same semantics as the CVE-2026-49855 fix. (pycurl.MAXFILESIZE alone is insufficient: it applies to the compressed transfer size.)

Affected versions

• <= 6.5.8 (latest tag; the curl client has never had a response-size limit) and master (6.6.dev1, verified 2026-08-15/17).
• 6.5.6's CVE-2026-49855 fix covered SimpleAsyncHTTPClient/http1connection.py only; curl_httpclient.py was not touched.

Credit

Reported by the diff/ambidiff security research effort (afldl).