Authorized clients

[Martin Peres]

On 09/01/2014 22:16, Jasper St. Pierre wrote:
> On Thu, Jan 9, 2014 at 3:54 PM, Martin Peres <martin.peres at free.fr 
> <mailto:martin.peres at free.fr>> wrote:
>
>     Le 09/01/2014 20:36, Jasper St. Pierre a écrit :
>
>         Security is based on trust. If you trust nothing, the computer
>         won't do very much for you. You can't even trust it to compute
>         correctly.
>
>
>     Security is based on access control. Every program that exposes a
>     service should think about how can it be abused and should
>     implement protocols as inherently-safe as possible. It should also
>     implement mitigations when it is impossible to propose the service
>     otherwise (an authentication system limiting the number of
>     attempts per minute for instance).
>
>     However, I do agree we need to trust the program that exports the
>     service, of course ;)
>
>
>         What's the threat model here? Let's say that we design a
>         screenshot API, and in your opinion it's the worst thing ever.
>         Who's attacking? How are they attacking? How will you mitigate it?
>
>
>     Good, some methodology! I already answered it earlier but whatever :)
>
>
> Sorry, I must have missed it. It's a large thread.
No probs. I understand. We'll need to make a recap email soon :)

>     The worse thing that can happen is an application running with the
>     user's uid grabbing and sending periodical screenshots to a
>     distant server running OCR and waiting for you to enter your bank
>     details on amazon.com <http://amazon.com>. As for how this
>     application got installed in the first plase, do I really have to
>     list all the ways or can we both agree this is out of scope?
>
>
> I feel it's out-of-scope from the point of view of Wayland. If there's 
> an application on your computer that is trusted to handle screenshots, 
> I don't believe it's the job of the Wayland compositor to second-guess 
> it.t into

Ask yourself, trusted by whom?

You are partly right, the protocol should kind of be agnostic of the way 
the authentication works. On the other hand, I'm kind of leaning towards 
thinking the screenshot protocol almost shouldn't exist (the solution 
[1] I talked about before). However, we'll need a video capture protocol 
and I think we all agree on how to handle the authentication on this one 
because we just can't limit the number of captures :D

>
> It's not Wayland's problem if the NSA silently planted a malicious, 
> trusted application on your computer. It's certainly dangerous that 
> this could happen, paranoia doesn't make for a good user experience: 
> imagine if your computer was a crazed conspiracy theorist :)

The NSA may find a way, but when a script kiddy can do it, then there is 
obviously something wrong. So, unless you are the kind of guys who 
authenticate over the internet without SSL, then you need to revise your 
judgement on that ;)
>
>     Who is attacking:
>     - Someone interested in what you are doing (industrial espionage,
>     justice department, government/NSA, spouse)
>     - Someone interested in your credentials (bank account, company
>     logins, etc...)
>     - Fun? (Yeah, some people have nothing better to do with their lives).
>
>     How are they attacking: Well, it depends on how (un)secure your
>     system is. I cannot
>
>     How do I mitigate this? The real solution would be not to have
>     screenshots at all. However, we want to provide screenshots to the
>     user if that's what he wants (responsibility on him though) so we
>     need to make sure the user is indeed the one that is willing to
>     get a screenshot. As we cannot trust applications, the server
>     needs data to prove this is indeed the user who wants this
>     screenshot. I see 3 possibilities here:
>     - He/she pressed a key that is meant for that
>     - He/she accepts that the app is trying to access the screenshot
>     functionnality (pop up)
>     - He/she disabled security until further notice or reboot (pekka's
>     idea, no permanent setting)
>
>     So how do we mitigate? By making sure no app is doing screenshots
>     in the user's back. EVER (unless explicitly asked by him, but that
>     should not be a permanent decision).
>
>
>         In our threat model of sandboxed applications, which takes
>         inspiration from Android, apps are given a coarse-grain set of
>         permissions: this app can take screenshots. That app can
>         access your contact list.
>
>         These are guarantees that the user can see, and the system can
>         provide. If the Flashlight app requires bluetooth, then
>         something sketchy might be going on, and the user can "vote
>         with their feet" and install a separate app.
>
>
>     I'm a big fan of these permission-set. I see them as the only way
>     for a user to know what is going on. The permission set is a
>     contract that a user either accepts or refuses. However, and we
>     need to be very clear about this, no other application should be
>     able to (ab)use the right of this application unless there is a
>     direct consent of the user (and this is getting sketchy).
>
>
> Of course not. It would be bad if somebody managed to repeatedly call 
> the trusted screenshot application to silently dump it somewhere. A 
> well-designed screenshot application shouldn't silently take a 
> screenshot and put it on the clipboard -- it would flash the screen, 
> make a camera noise, and maybe bring up a dialog for where to save it 
> to. I'd imagine that some people might build "take a screenshot 
> silently" scripts for their convenience, but if they want this I feel 
> that they have to deal with the consequences.

See, we agree on almost everything. The only thing we disagree about it 
the "trusted" label. At first, you trust the app, but then, there is a 
new version that allows disabling the visual cues and settings are 
stored in a file owned by the user. Then an attacker simply has to 
change the settings, run the app in the background and get the images. 
So, as I said, I don't trust application developers for doing the right 
things!

You'll say it is the same problem for the video protocol, but the 
difference is we don't have a choice.
>
> Apps on Android don't have this problem because there's a very limited 
> API to interact with other apps. I'd love this too, but I feel like 
> the Linux-using crowd wouldn't.
>
> Again, I don't feel it's the compositor's responsibility to protect 
> the user from trusted apps.
>
>         Locking down permissions UAC-style does not provide additional
>         security. Again, when the user is accomplishing a task, and
>         you give them a security prompt, you will get answers worse
>         than random chance. The user is doing something, now they have
>         to make a decision about yes/no. They are often ill-equipped
>         to handle such questions.
>
>
>     UAC is a piece of shit. Even as a security engineer, I cannot make
>     a good decision because they just basically say "I need more
>     privilege, do you agree?".
>     Because users don't seem to care anymore, the little security
>     benefit it could have renders the whole system useless. UAC is
>     just proving that an MLS security policy ultimately resolves into
>     running everything into the highest privileged mode.
>
>
>         If they press a "Take Screenshot" button in the app, and the
>         user is now prompted with a "Are you sure you meant to take a
>         screenshot?" prompt, what happens if they hit No to get rid of
>         the annoying prompt? What does the app do? What does the user do?
>
>     What happens if the user hits no: Simple, you get no screenshot.
>     When you visit a website with chrome and the website wants to know
>     your location, you say yes or no. If the website cannot deal with
>     it, it can't.
>
>
> So, in the most common case, the app breaks. If the user clicked on 
> the "Take Screenshot" button, and then picks a button from the prompt 
> at random, they get a broken app 50% of the time when they wanted to 
> take a screenshot. I'd say this isn't a worthwhile user experience.

If the user doesn't want the screenshot, the app won't get the 
screenshot. How is this broken?
>
>     From your previous email, you seem to be interested in UDAC. This
>     is a paper I haven't taken the time to read yet, but if I'm right,
>     it requires a way for applications to import "secure" widgets into
>     their windows. In this case, there could indeed be a
>     compositor-controlled "take screenshot" button and the application
>     couldn't change it's meaning. While it is something that could
>     eventually be done in Wayland using sub-surfaces, this is not
>     doable right now so there is no point discussing it in here. If
>     you want to discuss about it, we are a group of 3 security
>     engineers and one is doing his PhD on usable security. You may be
>     interested in joining us (if that's ok with the others).
>
>     If we manage to make an implementation of UDAC in wayland, then we
>     would be able to implement a screenshot application like you
>     wished to have. However, it won't be portable and will basically
>     be an equivalent to the compositor running an application with a
>     custom protocol to request whatever screenshot it wants.
>
>     Without those secure widget, the compositor doesn't know that a
>     click on the mouse at this position actually means "take a
>     screenshot".
>
>
> I tried to Google for "UDAC" and "UDAC security", but couldn't find 
> anything about this. This sounds to me a bit like Portals -- instead 
> of giving the app free roam to browse the user's documents, the app 
> asks the OS to let the user pick a file from its documents, and gets 
> passed back file contents. Apps can't access the user's sensitive data 
> directly by scanning, only the files that the user chooses to share.

That's the idea. I haven't read the paper yet, but I know it is much 
more than that.
>
> I'd hope that we'd be able to have such technology with Wayland, given 
> how it's technology we're working on for application sandboxing, and 
> half the reason kdbus is being developed.

Yep, but this is out of scope. The thread is already long-enough.