[jfager]

I don't think I understand what you're objecting to. The way I read it, CSP is lock-down by default (if the header is sent across to opt into it), with the application developer being responsible for providing a whitelist of where code and data can be loaded from. How can you do something similar today without still taking care to escape all inputs, etc.?

http://people.mozilla.org/~bsterne/content-security-policy/d...

[tptacek]

The libraries people use to "escape all inputs, etc" are providing effectively the same functionality as CSP is, but that's not my real concern.

My real concern is, despite the fact that developers have the ability to set policies about what regions on the page can contain dynamic content, "policy" is generally too brittle to describe what people need to put on pages in real-world apps.

[jerf]

I don't see anything in the spec about setting "regions" of dynamic content: https://wiki.mozilla.org/Security/CSP/Spec

I understand how you could think that was what was involved, as I got the same impression from the article and also thought that was dumb. However, that's not what is going on. You shut down broad classes of functionality entirely, and are required to provide all legit Javascript through <script> tags, which furthermore must come from your whitelisted sources.

(Actually, there are some other modes too; consult the spec for details. However, I expect this is the one people will be talking about.)

The article isn't exactly wrong, but it doesn't accurately convey what is going on, either. Basically, if you can discipline yourself well enough to only use included files (which isn't a terrible style anyhow), then this allows you to ensure that you won't execute injected content.

Is this something that you should be correctly escaping? Yes, absolutely. I am a huge advocate of that in my workplace. But I would certainly take advantage of this extra layer of protection, too, rather than rely on everybody in my company getting it all absolutely right, all the time. The problem with something like escaping is that you make one little error and you lose everything. This goes a long way towards mitigating that.

I haven't done a lot of thinking about this, but it looks good; in general, voluntarily discarding privileges you don't intend to use is a good security practice.

[tptacek]

It's my fault for being imprecise, but I'm saying, developers can already designate regions of their output as "XSS-safe" or "script-free", without a browser extension.

[jerf]

Actually, I tend to agree with your point in the other comment you made in this thread, asking who will actually use it. As one of the few who care in my company, I am both the only one who might set this up, and one of the few who are careful with my encoding in the first place. It does seem an awfully narrow target of people who can't/won't encode correctly, but will do this thing which will also be hard.

Still, I might use it, contingent on broader support and some burn-in time to ensure that it doesn't somehow create some sort of huge hole itself.

[jfager]

The libraries people use to "escape all inputs, etc" are providing effectively the same functionality as CSP is

Maybe, but not everyone's using those libraries, and those libraries may themselves have bugs or may be used incorrectly. Defense-in-depth, and all that.

"policy" is generally too brittle to describe what people need to put on pages in real-world apps

Just reading through the spec I linked to, I don't see what would be particularly brittle about it. Do you have an example of a real-world use case that this would hamper?

[tptacek]

So, getting back to what I don't get about CSP, I don't see how anyone's more likely to use a new policy specification language better than they are going to use a toolkit designed for their own web framework.

Brittleness: CSP disables <script> tags and all the script element attributes. Almost every major website in the world uses these in some form today. Both CSP and "anti-XSS" features do server-side policy about what parts of a page can have scripts, but the "anti-XSS" features are far more flexible.

[jfager]

I don't see how anyone's more likely to use a new policy specification language better than they are going to use a toolkit designed for their own web framework.

This can be an http header, so an admin or senior can configure the whitelist once at the server layer and everyone in the org reaps the benefit, even if your development toolkit has problems of its own.

CSP disables <script> tags and all the script element attributes.

I don't do much front-end work, so correct me if I'm way off, but I was under the impression that the only reason to use script elements today was to either load your own scripts or to hack around the same-origin policy to load 3rd party scripts. That use case still works under CSP, as long as the site you're loading from is in the whitelist.

[tptacek]

That's a really naive view of how app development and deployment works, so I'll sum my response to it up with, "the admin who adds the header that breaks the Javascript in a major application is going to be fired within a single digit number of Unix timeslices of the change going live."

The problem with this stuff is that it requires virtually top-to-bottom surgery on real world apps. It is another naive view of applications that says that the dangerous inputs all come from form arguments.

[jfager]

Wait, what? Of course an admin wouldn't just turn it on out of the blue one day, and of course that would be really, really stupid, but admins on the edge are responsible for stuff like this all the time. Turning on compression, setting up firewall rules, configuring caches.

Companies serve old applications, companies serve applications built on different platforms, companies get caught without the resources to rewrite insecure applications that pay the bills. You really think this is something that an ops team wouldn't love to have in its toolbox, to use as needed and where appropriate?

The problem with this stuff is that it requires virtually top-to-bottom surgery on real world apps.

If you can't use it in your app or environment without a huge rewrite, then don't. If you can, then it's gravy.

It is another naive view of applications that says that the dangerous inputs all come from form arguments.

I never said or implied they did, and that's not the only threat this proposal addresses.