Repository navigation
Description
I'm Claude Opus 5.5, AI can make mistakes. Written for a human user, who asked for this report.
Filed as a public issue tagged security, which SECURITY.md allows as an alternative to emailing security@getstream.io.
Summary
verify_and_parse_webhook decompresses a gzip body completely before it checks X-Signature. The order is forced: the HMAC covers the uncompressed JSON. But gunzip_payload has no limit on how much it decompresses, so anyone who can reach a webhook endpoint can make it use hundreds of MB to GBs of memory with a small unsigned request, a classic "gzip bomb". A webhook endpoint has to be public for Stream to reach it, and the signature check only fails after the memory is already used.
Affected code
- stream-chat-python
stream_chat/webhook.py:gunzip_payloadcallsgzip.decompress(raw)with no limit, andverify_and_parse_webhookcalls it before_verify_and_parse.
stream-chat-python/stream_chat/webhook.py
Lines 44 to 58 in 24c93b2
def gunzip_payload(body: _BytesLike) -> bytes: """Return ``body`` unchanged unless it starts with the gzip magic (``1f 8b``, per RFC 1952), in which case the gzip stream is decompressed. Magic-byte detection (rather than relying on a header) means the same handler stays correct when middleware - Rails, Django, Laravel, Phoenix - auto-decompresses the request before your code sees it. """ raw = _to_bytes(body) if raw[:2] != GZIP_MAGIC: return raw try: return gzip.decompress(raw) except (gzip.BadGzipFile, OSError, EOFError) as exc: raise InvalidWebhookError(INVALID_WEBHOOK_GZIP_FAILED) from exc - stream-py (
getstream)getstream/webhook.py:gunzip_payloadcallsgzip.decompress(bytes(body)), used byverify_and_parse_webhook.
https://github.com/GetStream/stream-py/blob/f0f1da5661bd9a4ba131c5dd6e34f9c79151f987/getstream/webhook.py#L654-L678 - stream-chat-js
src/signing.ts:gunzipPayloadcallszlib.gunzipSync(body)withoutmaxOutputLength, used byverifyAndParseWebhook.
https://github.com/GetStream/stream-chat-js/blob/3c1a18a0be4d5086cce007b9e98ebab435bf386a/src/signing.ts#L162-L174
The SQS and SNS parsers (parse_sqs / parse_sns, parseSqs / parseSns) use the same helper. Their input normally comes from a queue we control, so the exposure there is lower.
Reproduction (stream-chat 4.32.0, Python 3.14)
import tracemalloc, zlib
from stream_chat.webhook import verify_and_parse_webhook
from stream_chat.base.exceptions import InvalidWebhookError
def bomb(size: int) -> bytes: # gzip of `size` zero bytes, built without holding `size` bytes
d = zlib.compressobj(9, zlib.DEFLATED, 31)
chunk = bytes(1 << 20)
return b"".join([d.compress(chunk) for _ in range(size >> 20)] + [d.flush()])
body = bomb(256 << 20) # 260,934 bytes on the wire
tracemalloc.start()
try:
verify_and_parse_webhook(body, "not-a-signature", "secret")
except InvalidWebhookError as e:
print(e) # "signature mismatch", raised only after inflating
print(tracemalloc.get_traced_memory()[1] >> 20, "MB peak") # 557 MBA 255 KB unsigned request drives peak memory to ~558 MB before the signature is rejected. A ~1 MB request inflates to ~1 GB, and a handful of concurrent requests will run a typical container out of memory.
Suggested fix
Bound the output of decompression, and reject anything over the limit as an invalid webhook, so it can never succeed and costs at most the limit in memory:
- Python:
zlib.decompressobj(wbits=16 + zlib.MAX_WBITS).decompress(raw, max_length=LIMIT + 1). Treat output longer thanLIMITas an error, and check.eof(truncated stream) and.unused_data(trailing data). - Node:
zlib.gunzipSync(body, { maxOutputLength: LIMIT }), which throwsRangeErrorpast the limit. - The limit: a default sized well above any real event, overridable by callers. Stream already caps SQS messages at 256 KB, which may be a natural default. The recorded real chat webhook events we hold are 7–18 KB.
Documenting the limit in the webhook docs' compression section would also help people who decompress in their own middleware.
Context
We hit this while adding gzip support to our webhook receiver. We decompress through a capped zlib.decompressobj there instead of the SDK helper, and would happily switch back once the helper is bounded.
Activity