aiohttp: Unread Compressed Request Bodies Bypass client_max_size During Cleanup
Description
During cleanup, a compressed request body may be decompressed into memory in one chunk, allowing an attacker to trigger a DoS via a zip bomb edge case.
AI Insight
LLM-synthesized narrative grounded in this CVE's description and references.
During cleanup, a compressed request body may be decompressed into memory in one chunk, allowing an attacker to trigger a DoS via a zip bomb edge case.
Vulnerability
During cleanup of a compressed request body in aiohttp versions ≤3.14.0, the body may be decompressed into memory as a single chunk rather than in streaming fashion. This occurs because the code path handling cleanup does not properly limit decompression size for compressed request bodies [1][2].
Exploitation
An attacker with the ability to send HTTP requests to an affected server can craft a compressed payload (e.g., a zip bomb) that, when decompressed during cleanup, expands to a large amount of data. No authentication or special privileges are required; the vulnerability can be triggered remotely via a standard HTTP request containing a compressed body [1][2].
Impact
Successful exploitation leads to excessive memory consumption, potentially causing a denial of service (DoS) condition. This is described as a zip bomb edge case [1][2].
Mitigation
The vulnerability is fixed in commit 4f7480e, which ensures that compressed request bodies are streamed during cleanup. Users should upgrade to a patched version (e.g., >3.14.0). If unable to upgrade, disable compression as a workaround [1][2].
AI Insight generated on Jun 15, 2026. Synthesized from this CVE's description and the cited reference URLs; citations are validated against the source bundle.
Affected products
2Patches
14f7480e474cc[PR #12828/13b635d7 backport][3.14] Bounded unread compressed drain (#12845)
4 files changed · +104 −5
aiohttp/streams.py+12 −5 modified@@ -560,14 +560,21 @@ def _read_nowait(self, n: int) -> bytes: """Read not more than n bytes, or whole buffer if n == -1""" self._timer.assert_timeout() - chunks = [] + if n == -1: + # Drain only chunks present now; _read_nowait_chunk() can + # re-entrantly resume_reading() and refill the buffer. + count = len(self._buffer) + if count == 1: + return self._read_nowait_chunk(-1) + return b"".join([self._read_nowait_chunk(-1) for _ in range(count)]) + + chunks: list[bytes] = [] while self._buffer: chunk = self._read_nowait_chunk(n) chunks.append(chunk) - if n != -1: - n -= len(chunk) - if n == 0: - break + n -= len(chunk) + if n == 0: + break return b"".join(chunks) if chunks else b""
CHANGES/12828.bugfix.rst+1 −0 added@@ -0,0 +1 @@ +Fixed :meth:`~aiohttp.StreamReader.readany` and :meth:`~aiohttp.StreamReader.read_nowait` joining data fed back into the buffer during the call (when draining below the low water mark resumes reading) into a single unbounded :class:`bytes`; a call now returns only the chunks that were buffered when it started, keeping the drain of an unread auto-decompressed request body bounded by the read buffer -- by :user:`bdraco`.
tests/test_streams.py+28 −0 modified@@ -1723,3 +1723,31 @@ def resume_reading() -> None: protocol.resume_reading.assert_called() assert protocol._reading_paused is False + + +async def test_readany_does_not_drain_reentrant_refill( + protocol: mock.Mock, +) -> None: + """A single readany() must not reassemble data fed re-entrantly. + + Draining below the low water mark resumes reading, which can synchronously + refill the buffer (e.g. decompressing another chunk). Joining that refill in + one call would reassemble an unbounded body. + """ + loop = asyncio.get_running_loop() + stream = streams.StreamReader(protocol, limit=4, loop=loop) + + refills = [b"second", b"third"] + + def resume_reading() -> None: + if refills: + stream.feed_data(refills.pop(0)) + + protocol.resume_reading.side_effect = resume_reading + + stream.feed_data(b"first") + + # Popping "first" refills "second", but this readany() returns only "first". + assert await stream.readany() == b"first" + assert await stream.readany() == b"second" + assert await stream.readany() == b"third"
tests/test_web_functional.py+63 −0 modified@@ -4,7 +4,9 @@ import pathlib import socket import sys +import zlib from collections.abc import Generator +from contextlib import suppress from typing import Any, NoReturn from unittest import mock @@ -24,7 +26,9 @@ ) from aiohttp.compression_utils import ZLibBackend, ZLibCompressObjProtocol from aiohttp.hdrs import CONTENT_LENGTH, CONTENT_TYPE, TRANSFER_ENCODING +from aiohttp.helpers import DEFAULT_CHUNK_SIZE from aiohttp.pytest_plugin import AiohttpClient, AiohttpServer +from aiohttp.streams import StreamReader from aiohttp.typedefs import Handler from aiohttp.web_protocol import RequestHandler @@ -1711,6 +1715,65 @@ async def handler(request): await resp.release() +@pytest.mark.parametrize("decompressed_size", [4 * 1024 * 1024, 32 * 1024 * 1024]) +async def test_unread_compressed_body_drain_is_bounded( + aiohttp_server: AiohttpServer, + monkeypatch: pytest.MonkeyPatch, + decompressed_size: int, +) -> None: + """Draining an unread compressed body stays bounded by the read buffer. + + A handler that rejects before reading still drains the payload during + lingering close; a small compressed body must not force a large transient + allocation (a deflate-bomb style DoS). + """ + drain_reads: list[int] = [] + drained = asyncio.Event() + readany = StreamReader.readany + + async def record_readany(self: StreamReader) -> bytes: + data = await readany(self) + assert data + drain_reads.append(len(data)) + drained.set() + return data + + monkeypatch.setattr(StreamReader, "readany", record_readany) + + async def handler(request: web.Request) -> web.Response: + return web.Response(status=401) + + app = web.Application(client_max_size=1024) + app.router.add_post("/", handler) + server = await aiohttp_server(app) + + body = zlib.compress(b"a" * decompressed_size) + assert len(body) < decompressed_size + head = ( + b"POST / HTTP/1.1\r\n" + b"Host: localhost\r\n" + b"Content-Encoding: deflate\r\n" + b"Content-Length: %d\r\n" + b"Connection: keep-alive\r\n\r\n" + ) % len(body) + + reader, writer = await asyncio.open_connection(server.host, server.port) + try: + writer.write(head + body) + await writer.drain() + status_line = await asyncio.wait_for(reader.readline(), 5) + assert status_line.startswith(b"HTTP/1.1 401 ") + await asyncio.wait_for(drained.wait(), 5) + finally: + writer.close() + with suppress(ConnectionResetError, BrokenPipeError): + await writer.wait_closed() + + # Bounded by the buffer, not the decompressed size. + assert max(drain_reads) <= 3 * DEFAULT_CHUNK_SIZE + assert max(drain_reads) < decompressed_size + + async def test_app_max_client_size(aiohttp_client) -> None: async def handler(request): await request.post()
Vulnerability mechanics
Synthesis attempt was rejected by the grounding validator. Re-run pending.
References
2News mentions
1- Aiohttp: Nine CVEs Disclosed in a Single Day, Five Memory-Exhaustion DoS FlawsVypr Intelligence · Jun 15, 2026