Re: [RFC] kconfig: add hardened defconfig helpers

[Salvatore Mesoraca]

2018-07-20 15:37 GMT+02:00 Eric W. Biederman <ebiederm@xxxxxxxxxxxx>:
> Kees Cook <keescook@xxxxxxxxxxxx> writes:
>
>>> +CONFIG_USER_NS=n
>>> +~~~~~~~~~~~~~~~~
>>> +
>>> +**Negative side effects level:** Medium
>>
>> Unfortunately I think this is High or Extreme. USER_NS gets a lot of use.
>>
>>> +**- Protection type:** Attack surface reduction
>>> +
>>> +This allows containers to use user namespaces to provide different
>>> +user info for different servers.
>>> +User namespaces have been abused in the past for privilege
>>> +escalation.
>
> This is not a particularly good description.

You are right, I've been a bit shallow...
Thank you for pointing it out.
I didn't mean to say that user namespaces are bad for security in general.
I'll make this less ambiguous in the next revision.

> User namespaces do indeed increase the attack surface of programs that
> don't use them.
>
> User namespaces when used to build ``unprivileged containers'' remove or
> at least drastically reduce the need to run as root when setting up
> containers.  Which is attack surface reduction.
>
> User namespaces make available tools that are commonly used to build
> sandboxes.  Chrome for example uses users namespaces if they are
> available as part of setting up it's sandbox.
>
> User namespaces are not limited to containers.
>
> The bugs in the attack surface that user namespaces expose that have
> been used for privilege escalation have to the best of my knowledge
> closed.  So while there is some danger in the increased attack surface
> we are looking at implementation defects rather than design defects.
>
> Eric