Impact
In runc, netlink is used internally as a serialization system for specifying the relevant container configuration to the C portion of our code (responsible for the based namespace setup of containers). In all versions of runc prior to 1.0.3, the encoder did not handle the possibility of an integer overflow in the 16-bit length field for the byte array attribute type, meaning that a large enough malicious byte array attribute could result in the length overflowing and the attribute contents being parsed as netlink messages for container configuration.
This vulnerability requires the attacker to have some control over the configuration of the container and would allow the attacker to bypass the namespace restrictions of the container by simply adding their own netlink payload which disables all namespaces.
Prior to 9c44407, in practice it was fairly difficult to specify an arbitrary-length netlink message with most container runtimes. The only user-controlled byte array was the namespace paths attributes which can be specified in runc's config.json, but as far as we can tell no container runtime gives raw access to that configuration setting -- and having raw access to that setting would allow the attacker to disable namespace protections entirely anyway (setting them to /proc/1/ns/... for instance). In addition, each namespace path is limited to 4096 bytes (with only 7 namespaces supported by runc at the moment) meaning that even with custom namespace paths it appears an attacker still cannot shove enough bytes into the netlink bytemsg in order to overflow the uint16 counter.
However, out of an abundance of caution (given how old this bug is) we decided to treat it as a potentially exploitable vulnerability with a low severity. After 9c44407 (which was not present in any release of runc prior to the discovery of this bug), all mount paths are included as a giant netlink message which means that this bug becomes significantly more exploitable in more reasonable threat scenarios.
The main users impacted are those who allow untrusted images with untrusted configurations to run on their machines (such as with shared cloud infrastructure), though as mentioned above it appears this bug was not practically exploitable on any released version of runc to date.
Patches
The patch for this is d72d057 and runc 1.0.3 was released with this bug fixed.
Workarounds
To the extent this is exploitable, disallowing untrusted namespace paths in container configuration should eliminate all practical ways of exploiting this bug. It should be noted that untrusted namespace paths would allow the attacker to disable namespace protections entirely even in the absence of this bug.
References
Credits
Thanks to Felix Wilhelm from Google Project Zero for discovering and reporting this vulnerability. In particular, the fact they found this vulnerability so quickly, before we made a 1.1 release of runc (which would've been vulnerable) was quite impressive.
For more information
If you have any questions or comments about this advisory:
felixwilhelm Analyst
Impact
In runc, netlink is used internally as a serialization system for specifying the relevant container configuration to the C portion of our code (responsible for the based namespace setup of containers). In all versions of runc prior to 1.0.3, the encoder did not handle the possibility of an integer overflow in the 16-bit length field for the byte array attribute type, meaning that a large enough malicious byte array attribute could result in the length overflowing and the attribute contents being parsed as netlink messages for container configuration.
This vulnerability requires the attacker to have some control over the configuration of the container and would allow the attacker to bypass the namespace restrictions of the container by simply adding their own netlink payload which disables all namespaces.
Prior to 9c44407, in practice it was fairly difficult to specify an arbitrary-length netlink message with most container runtimes. The only user-controlled byte array was the namespace paths attributes which can be specified in runc's
config.json, but as far as we can tell no container runtime gives raw access to that configuration setting -- and having raw access to that setting would allow the attacker to disable namespace protections entirely anyway (setting them to/proc/1/ns/...for instance). In addition, each namespace path is limited to 4096 bytes (with only 7 namespaces supported by runc at the moment) meaning that even with custom namespace paths it appears an attacker still cannot shove enough bytes into the netlink bytemsg in order to overflow the uint16 counter.However, out of an abundance of caution (given how old this bug is) we decided to treat it as a potentially exploitable vulnerability with a low severity. After 9c44407 (which was not present in any release of runc prior to the discovery of this bug), all mount paths are included as a giant netlink message which means that this bug becomes significantly more exploitable in more reasonable threat scenarios.
The main users impacted are those who allow untrusted images with untrusted configurations to run on their machines (such as with shared cloud infrastructure), though as mentioned above it appears this bug was not practically exploitable on any released version of runc to date.
Patches
The patch for this is d72d057 and runc 1.0.3 was released with this bug fixed.
Workarounds
To the extent this is exploitable, disallowing untrusted namespace paths in container configuration should eliminate all practical ways of exploiting this bug. It should be noted that untrusted namespace paths would allow the attacker to disable namespace protections entirely even in the absence of this bug.
References
Credits
Thanks to Felix Wilhelm from Google Project Zero for discovering and reporting this vulnerability. In particular, the fact they found this vulnerability so quickly, before we made a 1.1 release of runc (which would've been vulnerable) was quite impressive.
For more information
If you have any questions or comments about this advisory: