I get to live in a beautiful world. The world that I get to live in is one in which everything works, except for the things that I am actually responsible for.
I assume that the rest of the Chrome security team is doing their job. I assume that the CPU is actually doing what it's supposed to be doing on a regular basis. And in this beautiful and completely hypothetical world, we still have issues, and I'd like to talk with you about some of them today.
The painting behind me is Ulysses and the Sirens. Are any of you familiar with that painting? Culture. Excellent.
The sirens sing beautifully. They sing so beautifully in fact that it drives men to the brink of madness. So when a sailor goes by their island the sailor will listen. They'll hear the song of the sirens, all they'll want to do is hear more of that song, so they'll throw themselves off the side of the ship, in order to get to the island to get more and more of this beautiful music. And then the sirens will eat them because, you know, that's what sirens do.
Ulysses is a man who is out for the Trojan War, and he had been away from home for, like, ten years. And all he wants to do at this point is get back home with his men.
He talks to a witch named Circe, and she tells him that to get back home he's going to have to go by this island. There's simply no other way, so how can he defend himself against these sirens?
As it turns out, Circe has a good idea for him. You can take wax. You can mold the wax into little earplugs, and you can put the wax in your ears. And then you don't hear the song of the siren, and by not hearing the song you're not driven mad. You don't jump overboard, and you're not eaten. It sounds like a great scenario.
Circe further tells him that if he wants he can actually listen to the song of the sirens by having his men bind him to the mast, right. So if you tie your hands and your feet tightly to the mast, then he can hear the song of the sirens, but he's physically prevented from actually jumping off the ship.
He thinks this sounds like a great plan, so he has his men tie him to the mast. They all put wax in their ears, and they start sailing by the island, and as it turns out this works beautifully.
Ulysses hears the song of the sirens. He hears this beautiful music. He wants to hear more of it, but he has physically prevented himself from actually jumping off the side of the ship.
Now, this is incredibly relevant to security on the web. Let's look at the threat model that we actually care about, in order to see how that might fit in.
So there are basically two things that I really care about with regard to the web platform. Assuming we implement all of the specs correctly there are two things that really worry me.
The first is cross-site scripting. Scott talked about this a little bit earlier today. I want to give you a little bit more detail, just to make sure we're all on the same page with regard to the dangers of content injection, and then we'll move on.
Behind me you see five different places in which content can be injected into pages, and all of these five places have different rules. So if you're injecting something into a style sheet, you're going to want to escape that differently than if you're injecting it into a paragraph or into an HTML attribute. All of these have different rules, but these rules are completely mechanical. XSS is a completely solved problem.
The next kind of attack that I think is really interesting is something called
cross-site request forgery. The idea here is that the browser has ambient
authority for a wide variety of origins. If I'm logged into
innocentvictim.com, then all of the requests it makes and it expects will have
cookies associated with them and other sorts of authentication data. That means
innocentvictim.com will honor my requests and will actually take whatever
action it is that I'm asking for. If
evil.com however makes a sneaky request
on the side by loading an iframe, or loading an image, or sending an XHR
request, that request can have the same authentication tokens bound to it. What
this means is that it's difficult for
innocentvictim.com to distinguish
between good requests and bad requests, and it's entirely possible for a third
party to maliciously force that victim site to do something that it wasn't
Now I said cross-site scripting is a completely solved problem. It is theoretically, but practically what we see in the Vulnerability Rewards Program at Google, something that Fillipo talked about earlier today, the vast, vast majority of the attacks that we actually pay for are cross-site scripting. Even at Google. A company with a dedicated team doing nothing but preventing the developers inside of Google from doing whatever it is they want to be doing. We build really interesting tools, and we build out good, solid templating systems, and yet we still have cross-site scripting attacks.
So what can we do to start mitigating these? And why do they happen in the first place?
Well, as it turns out the browser is a confused deputy. That's the name of this class of attack. It has power, and it can act on that power, but by doing so it's not actually taking actions on your behalf. It's taking actions on an attacker's behalf using your credentials.
Now, you shouldn't give a deputy like this a gun. You should instead give them something a little bit softer.
The name of the game is privilege reduction. You heard earlier the notion of the principle of least privilege. If you don't have any privileges, then it doesn't matter if someone asks you to do something bad, because you have prevented yourself from doing that. You don't have the privilege to do so.
The first of these you've already heard about. It's called Content Security Policy, and the idea of Content Security Policy, as Scott mentioned earlier today, is to give you granular control over the things that are happening on your website. So he talked a little bit about a fairly simple set of Content Security Policy that can be applied to particular sites. Now, as you can see, this is a fairly simple content security policy that's used on a very small site called Twitter. This is copy and pasted this morning, and it's ridiculous, right? This thing is absolutely huge, and as it turns out it's not nearly as strong as you would expect from the amount of text that you see on the page.
The idea of course is that it wants to allow you to talk to Twitter, but it doesn't
want to allow you to talk to
evil.com. So load script, or load images, or any of
these things. What we've seen however is that these kinds of white lists are
actually really, really difficult to get right. They're hard to maintain, and if you
make them easy to maintain they're too loose to have any actual impact.
So the folks at Cure53 -- Mario Heiderich and a few others -- have noted that on
large origins, origins like
google.com or various CDNs, there's enough code there
to do absolutely anything you want to do.
So if you whitelist
google.com, you're not only whitelisting those individual
files that you yourself wrote. You're whitelisting a wide variety of things
including old versions of Angular that you actually do not ever want to run on a
There has to be a better way, and in CSP3 we're looking at a couple of options that the Google folks think are really interesting and I think are pretty interesting in the browser as well.
The idea is called
'unsafe-dynamic', and it treats nonces and hashes as
capability tokens. So you put a script tag on your page, and the script tag
contains a particular attribute. The attribute contains a long random token.
That token is reflected in the HTTP header that you sent to the page.
This means that you're not whitelisting any script. You're whitelisting this
particular script, and because the nonce is long and random, the attacker can't
actually guess it. This has some really good properties, but we can go one step
We can allow the script that you've loaded, the script that you've carefully audited, and the one that you know is doing the right thing to then load its own scripts. That is we'll use nonces for every script on the page or hashes for every script on the page, and what that does is allow us to say, "We've audited all of these scripts. We know these are the ones that ought to run," and then those scripts are empowered to do whatever it is that that widget needs to do. So if you want to load Google Maps for instance, that's going to load a script which loads a couple of more scripts and then shows a map.
Now, this is possible to do today with whitelists. You can whitelist absolutely every file that you might need. That becomes very brittle, very difficult to maintain, or very easy for the library owner to actually break you, because the library owner doesn't know that you're using a very strict CSP and probably isn't going to notify you when they update it.
So we think unsafe dynamic is a really interesting property. It's going to be shipping in Chrome 52, and we'd really like people to start playing around with it, so that we can kind of understand the properties and understand whether it actually solves the problems that we think it's going to solve as well as we think it's going to solve them.
The next thing I'd like to talk about is called Suborigins. Now, all of you are probably familiar with the idea of the same origin policy. Every website on the Internet has a tuple of a scheme, a host, and a port. So "HTTP", "example.com", "80". That defines an area of influence and a set of data that everything executing within the context of an origin has access to.
Now, I'll repeat that. Everything in that origin has access to everything in the origin, so if I execute some script in example.com, then I get all the local storage. I can talk to the backend and pretend that I am example.com. This ends up being a good isolation property for most of the web, but for a lot of origins that have multiple applications running on them we'd like to be able to segregate things even more.
So if you think about google.com for instance, we have Docs, we have Maps, and we have Gmail, and a variety of other applications including marketing applications from 2004 that are all running on google.com.
And those have very different levels of audit that have been applied to them, but because they're all running on google.com they all have that same set of privilege.
What we'd like to do is allow developers to actually shard that physical origin into a number of logical origins, so that you can say that, "This page that I'm loading is actually part of the Docs application, and this page is part of the Maps application." And they shouldn't be able to talk to each other directly. They shouldn't be able to access each other's data.
This is similar in concept to the notion of a sandbox, so iframes have this sandbox attribute that lets you push them into a unique origin.
The nice property here is that this is actually a named sandbox which means there are some properties about communication that we can start to enable, because we know that this is going to be the same name going forward.
Now, there are some questions about what exactly this sharding should amount to. What kinds of communications can we enable between the physical origin and these logical origins? What do we do with cookies? Because of course cookies span across origins. They span across domains, and they are a big pain in the butt for basically anyone who needs to deal with security on the web. So there are some open questions there, and it would be really helpful to us if you all would take a look at the prototype implementation that's in Chrome today and the specification that I showed just a moment ago, and help us figure out what kinds of use cases we should be addressing with this kind of functionality.
This is all experimental, and it's not something that I think we're going to ship in the very near future.
Subresource integrity, on the other hand, has already shipped in both Chrome and Firefox, and other browsers are starting to get interested in it. The idea is that you can whitelist a particular set of content coming from a CDN or coming from some external server. You can say, "I want this script to run, and not just this script at this location, but this exact script."
You put a digest into an
integrity attribute within the tag, and that gives you
the ability to say, "Not only this script from this location is going to execute,
but that we can verify that it is exactly the scrip that I think it's going to be."
This removes some of the potential for damage that CDNs can cause, because of course CDNs are a juicy target for attackers. If I can own CloudFlare, then I'm not only owning CloudFlare. I'm actually able to attack a wide variety of sites on the Internet.
By using an integrity attribute to get out to those CDNs, I can verify that the CDN is doing its job and that it hasn't been compromised. This has some really interesting properties.
A future version of this will probably do something more than just digests. I think we're very interested in looking at signatures. We're very interested in looking at the kinds of things that Diego talked about earlier. We're just not there yet with regard to this specification.
Finally, cookies. Cookies are super interesting. They're super interesting because they're super broken, from a security perspective. Cookies don't honor the same origin policy. They have no understanding of ports, and they span across schemes. This is kind of problematic.
We have a couple of ideas of how we can improve the state of cookies, so the HTTP working group in the IETF is going through some of these ideas, in order to figure out what pieces we can actually change without making breaking changes to the Internet. One idea that I think is pretty interesting is the notion of a same-site cookie. So I showed you earlier this picture of innocentvictim and evil.com, and evil.com is able to talk directly to innocentvictim because again the browser is an ambient authority.
It would be nice if innocentvictim could say, "You know, I don't actually want this scenario to ever happen. No one should ever embed me. I should never be in a situation where someone else has authority over me." So we can start looking at the initiator of a request and not just the recipient of a request when determining what cookies ought to be sent along with that request.
The idea is simple. The victim will specify a new attribute on the cookie called "SameSite." They can set it either as "strict" or "lax", and those have slightly different properties, but for the purposes today they're similar.
When evil.com makes a request -- maybe they've embedded an image, maybe they've embedded an iframe -- that request will be coming from something different than the victim site, and in that case because they've specified that this cookie should never be sent in that kind of scenario, we can actually prevent a large class of cross-site request forgery attacks. We think this is going to be really quite interesting for a lot of sites out there, and I think it has a good amount of potential to really reduce the impact of both cross-site request forgery attacks and cross-site script inclusion attacks which rely on the same sorts of principles.
Finally, I'd like to talk a little bit about TLS. So TLS is a little bit outside the model that we've been talking about before. It's not strictly about privilege reduction. It is instead about bringing up a baseline of security for your websites.
It is simply the case that if you're not delivering your websites over HTTPS, then you have no security. There is literally nothing you can do on the web, in browsers, that gives you any piece of security at all, if you don't have TLS as a baseline. TLS gives you a couple of guarantees. It gives you a modicum of confidentiality. It gives you integrity, because you know that the bits that the server sent actually are the ones getting to you. And it gives you authentication, because you know that that server is the one sending something to me, not any router in between me and that server.
The Chromium project has started to lock down old APIs that were initially shipped in such a way that they would work both over TLS and over non-TLS. Some examples of this are geolocation, your microphone, and your camera. Powerful features that we really never should have shipped over HTTP to begin with, because by shipping them over HTTP you are not just granting permission to the site you think you're granting permission to. You're actually granting permission to anyone between you and that site. Anyone in a privileged network position, and that's something that we simply can't allow, so we're starting to lock down those old APIs.
We're also looking at ways of making migration a little bit simpler. Scott talked a little bit about mixed content before, and mixed content has come up very often as a hard blocker for folks moving over to HTTPS. There are some CSP-related mechanisms that you can use, like block-all-mixed-content and upgrade-insecure-requests.
One more thing that we're thinking about doing is actually reversing the order of HSTS and mixed content checks. That is, if you set a strict transport security header that says, "Only connect to me over HTTPS," we would like to do that upgrade before we do mixed content checks.
We can't do that today, because there's a bit of a determinism problem. That is, developers have always been to all of their origins, so they will never see an HSTS error if they embed something in their page that isn't using the right scheme. Users will see it. They'll file bugs. The developer will throw up their hands and say, "Well, it works for me," and close the bug.
So we'd like to ensure that we can start doing HSTS first, but do it safely. And the idea that we have is to send out what's called a priming request. If I'm about to request an insecure object or an insecure resource, I'll first go out to the website over HTTPS, check to see whether I get an HSTS header back, and if I do, then I can do the upgrade, and everything will work out pretty well. This should be really valuable for large sites, especially YouTube, because YouTube has just started sending an HSTS header which gives us the ability to not block folks from upgrading, because if they upgrade, then they couldn't get YouTube, and they'd be a little bit sad.
Privilege reduction is the name of the game. There are a number of mechanisms that have been released over the last six months to a year that are really, really interesting, and I'd love for y'all to go out, and play around with them, and give us some feedback. Both on the Chrome team, but also in standards bodies, so that we ensure that we're building the tools that allow you to lock down your sites and to tie it up a little bit because, you know, sites shouldn't jump overboard.
Thank you very much.