[cmake-developers] Security in CMake

[Roger Leigh]

CMake is not and can never be a replacement for a package manager.  They 
serve entirely different roles.  While you can extend CMake to give it 
the ability to build and install all sorts stuff, that does not make it 
a package manager.  It's a convenient way to build a complex set of 
interdependent packages, but that's all it is.  I think trying to make 
it do more than that is dangerous; it's currently got a fairly clearly 
defined scope, and there are numerous established tools to take care of 
packaging.  Basically, I do worry about the potential for scope creep here.

Speaking from my own experience as one of the main authors of the Debian 
build tools (sbuild, schroot):

- every package is built individually in a secure sandbox which is 
created before the build and deleted after it.  I wrote a tool 
specifically for this purpose (schroot)--predating docker, openstack and 
all the current hip stuff.  You should probably use a more up to date 
technology yourself; the main point I wanted to make was the isolation 
of the filesystem, network, user etc. for the build step; we run the 
build step without any functional networking other than the loopback to 
ensure every package cleanly builds as a self-contained entity

- a package might be using CMake for its build, inside the sandbox, but 
CMake has no involvement at a higher level

This means that any malicious or accidental privilege escalation, file 
deletion or whatever is contained and will not cause problems.  And if 
there is anything wrong, we blow the whole lot away as a matter of 
course anyway.

- when the packaged binary is installed, there is no code execution 
outside the (optional) maintainer scripts; everything else is just a 
static set of files which get installed.  This is completely decoupled 
from CMake, or whatever build system was used for the build step.

When it comes to CMake building using a super-build and/or running 
find_package and executing third-party Config scripts, other CMake 
modules etc. this is clearly unsafe.  There's clearly scope for 
unbounded code execution by this code.  While these might be required 
for the build step, there's no need for them to have any influence or 
involvement for the package creation step and subsequent 
installation/upgrade/removal by the package manager.


Regards,
Roger

On 20/08/16 22:42, Egor Pugin wrote:
> Hi Tobias,
>
> You are right at many points. It's hard to really secure the system
> from build system/build artifacts/3rd party apps.
> But step by step it's possible to decrease number of potential sources
> of vulnerabilities.
> Described cmake features would be very helpful for this.
> Right now I'm just investigating possible security improvements and
> this is only 'nice-to-have' feature.
>
> On 21 August 2016 at 00:25, Tobias Hunger <tobias.hunger at gmail.com> wrote:
>> Hi Egor,
>>
>> Am 20.08.2016 13:48 schrieb "Egor Pugin" <egor.pugin at gmail.com>:
>>>
>>> Hi,
>>>
>>> I'm working on a package manager based on cmake.
>>> And some cmake instructions are downloaded with user packages.
>>> I'd like to have an ability to deny some cmake features in such
>>> external untrusted insertions.
>>
>> I am no CMake expert, but you are talking about securing a program that is
>> meant to take arbitrary input and run user-defined commands on that to
>> produce possibly executable output.
>>
>> I do not see any safe subset of CMake commands that is still able to do
>> anything useful.
>>
>> I can see a way for "insertions" to be useful, that does not involve them
>> changing the configuration (e.g. for a cross compiler), involve running some
>> 3rd party program (e.g. to add support for a new documentation system,
>> parser generator or whatnot), or the production of build artifacts (e.g.
>> build some library for the developer to use).
>>
>> *All* of  these are inheritently unsafe.
>>
>> Configuration change: Change the C compiler to rm and pass force -rf -- / as
>> flags.
>>
>> 3rd party program: Run rm -rf / when some certain input file is seen.
>>
>> Build artifacts: Put running rm -rf / into the binary/library so that this
>> is run during normal development workflow.
>>
>> I would try to run my package manager in an environment where running rm -rf
>> is harmless to the overall system health. Virtual machines or containers
>> spring to mind there. Not sure that is feasible.
>>
>> Or come up with insertions signing, etc. so that users can at least know
>> they got what was uploaded and know whom to blame when their systems get
>> wiped.
>>
>> Beat Regards,
>> Tobias
>
>
>