CVE-2023-41337

h2o is an HTTP server with support for HTTP/1.x, HTTP/2 and HTTP/3. In
version 2.3.0-beta2 and prior, when h2o is configured to listen to multiple
addresses or ports with each of them using different backend servers
managed by multiple entities, a malicious backend entity that also has the
opportunity to observe or inject packets exchanged between the client and
h2o may misdirect HTTPS requests going to other backends and observe the
contents of that HTTPS request being sent. The attack involves a victim
client trying to resume a TLS connection and an attacker redirecting the
packets to a different address or port than that intended by the client.
The attacker must already have been configured by the administrator of h2o
to act as a backend to one of the addresses or ports that the h2o instance
listens to. Session IDs and tickets generated by h2o are not bound to
information specific to the server address, port, or the X.509 certificate,
and therefore it is possible for an attacker to force the victim connection
to wrongfully resume against a different server address or port on which
the same h2o instance is listening. Once a TLS session is misdirected to
resume to a server address / port that is configured to use an
attacker-controlled server as the backend, depending on the configuration,
HTTPS requests from the victim client may be forwarded to the attacker’s
server. An H2O instance is vulnerable to this attack only if the instance
is configured to listen to different addresses or ports using the listen
directive at the host level and the instance is configured to connect to
backend servers managed by multiple entities. A patch is available at
commit 35760540337a47e5150da0f4a66a609fad2ef0ab. As a workaround, one may
stop using using host-level listen directives in favor of global-level
ones.

Bugs

• <http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1059413&gt;