实际上，权威的 RFC，RFC 6455 The WebSocket Protocol，有一个解释。我在这里引用它：
10.3. Attacks On Infrastructure (Masking)
In addition to endpoints being the target of attacks via WebSockets,
other parts of web infrastructure, such as proxies, may be the
subject of an attack.
As this protocol was being developed, an experiment was conducted to
demonstrate a class of attacks on proxies that led to the poisoning
of caching proxies deployed in the wild [TALKING]. The general form
of the attack was to establish a connection to a server under the
"attacker's" control, perform an UPGRADE on the HTTP connection
similar to what the WebSocket Protocol does to establish a
connection, and subsequently send data over that UPGRADEd connection
that looked like a GET request for a specific known resource (which
in an attack would likely be something like a widely deployed script
for tracking hits or a resource on an ad-serving network). The
remote server would respond with something that looked like a
response to the fake GET request, and this response would be cached
by a nonzero percentage of deployed intermediaries, thus poisoning
the cache. The net effect of this attack would be that if a user
could be convinced to visit a website the attacker controlled, the
attacker could potentially poison the cache for that user and other
users behind the same cache and run malicious script on other
origins, compromising the web security model.
To avoid such attacks on deployed intermediaries, it is not
sufficient to prefix application-supplied data with framing that is
not compliant with HTTP, as it is not possible to exhaustively
discover and test that each nonconformant intermediary does not skip
such non-HTTP framing and act incorrectly on the frame payload.
Thus, the defense adopted is to mask all data from the client to the
server, so that the remote script (attacker) does not have control
over how the data being sent appears on the wire and thus cannot
construct a message that could be misinterpreted by an intermediary
as an HTTP request.
Clients MUST choose a new masking key for each frame, using an
algorithm that cannot be predicted by end applications that provide
data. For example, each masking could be drawn from a
cryptographically strong random number generator. If the same key is
used or a decipherable pattern exists for how the next key is chosen,
the attacker can send a message that, when masked, could appear to be
an HTTP request (by taking the message the attacker wishes to see on
the wire and masking it with the next masking key to be used, the
masking key will effectively unmask the data when the client applies
it).
It is also necessary that once the transmission of a frame from a
client has begun, the payload (application-supplied data) of that
frame must not be capable of being modified by the application.
Otherwise, an attacker could send a long frame where the initial data
was a known value (such as all zeros), compute the masking key being
used upon receipt of the first part of the data, and then modify the
data that is yet to be sent in the frame to appear as an HTTP request
when masked. (This is essentially the same problem described in the
previous paragraph with using a known or predictable masking key.)
If additional data is to be sent or data to be sent is somehow
changed, that new or changed data must be sent in a new frame and
thus with a new masking key. In short, once transmission of a frame
begins, the contents must not be modifiable by the remote script
(application).
The threat model being protected against is one in which the client
sends data that appears to be an HTTP request. As such, the channel
that needs to be masked is the data from the client to the server.
The data from the server to the client can be made to look like a
response, but to accomplish this request, the client must also be
able to forge a request. As such, it was not deemed necessary to
mask data in both directions (the data from the server to the client
is not masked).
Despite the protection provided by masking, non-compliant HTTP
proxies will still be vulnerable to poisoning attacks of this type by
clients and servers that do not apply masking.