Sentry: Handle Content-Security Policy (CSP) violation reports

[benvinegar]

Content-Security Policy (CSP) is a new browser feature that makes it possible to restrict JavaScript files and other assets to a trusted set of origins:

http://www.html5rocks.com/en/tutorials/security/content-security-policy/
http://www.w3.org/TR/CSP/

CSP has a reporting mechanism whereby violations can be automatically POSTed by the browser to a reporting endpoint. It would be awesome if Sentry supported CSP violation reports, making it super easy for Sentry users to identify and track CSP violations (e.g. possible XSS attacks).

More on CSP's reporting mechanism:

http://www.html5rocks.com/en/tutorials/security/content-security-policy/#reporting
http://www.w3.org/TR/CSP/#sample-violation-report

[dcramer]

One note: need to figure out how security is handled here (is there an Origin? Referrer?)

[christianvuerings]

+1

[xncoder]

+1

[oreoshake]

@dcramer there is a referrer, origin is null, cookies will be stripped unless allowed by CORS.

http://www.w3.org/TR/2013/WD-CSP11-20130604/#sample-violation-report

[joostdevries]

+1 a lot.

[hd-deman]

+1

[36degrees]

+1

[joostdevries]

anyone working on a pr for this?

[mattrobenolt]

@joostdevries I was, then realized that this is a mess.

• A browser can make _tons_ of requests to this, so we'd just be getting DDoS'd 24/7.
  
  
  
  • A separate HTTP request is made for every single violation. In my initial testing in my normal browser without touching anything, I put a strict CSP rule on a page and it made like, 12 HTTP requests to report the problems. One for every single browser plugin I had that tried to modify the page or something.
  
  
• How do we aggregate and report on this?
• How does this look in the UI?

Imagine every visitor to your site making 12+ error reports to Sentry on every page load. getsentry.com wouldn't be able to handle that and I doubt anyone trying to do this at any scale with lots of page views would die as well. I think there needs to be a lighter more specific mechanism for logging these. Treating them as a normal exception might not be the right approach here.

[alex]

@mattrobenolt it seems like they should be aggregated by {url, asset}, or maybe just {asset}; URLs might need to be turned into patterns somehow?

[alex]

(Also, anything that we come up with here probably makes sense for HTTP Public-key-pinning as well)

[dcramer]

@alex I think the main concern was just the volume of data. There doesn't seem to be a way to sample/restrict it, and JS events are already pretty high volume (I imagine this being higher).

This is more of a concern for getsentry.com, but until we have a really solid "you're going to pay us for the CPU you cost us" solution it's a bit scary to offer something like this. I see it in a similar vein as HTTP 404 logging.

[mattrobenolt]

Yeah, I think there's just a lack of information here so far. I should probably start tracking it somewhere just to collect data and see where we're at and what sense I can make of it. It's definitely a thing we should figure out though, I agree.

[mattrobenolt]

And yes, volume of data can be ridiculous, but I think can be solved.

[alex]

So, the primary use case I'm interested in is helping site maintainers detect resources that are inadvertantly accessed over HTTP, when they have a theoretically HTTPS-only site, to help them avoid mixed content warnings.

It seems like that use case can probably be acheived with less data. Are there other important use cases?

[mattrobenolt]

I think most of the use cases of CSP are valid. :)

The noisy/gross things are from browser plugins. Those are things I'd want to ignore. But since they don't go through a client, we don't have an opportunity to filter them out.

One of my ideas was to always relay through a raven-python or something similar to do filter on _your_ server first, and raven-python could expose a url to their Django app. But that's obviously very limited to raven-python and Django.

[alex]

it seems like expecting someone to have a server component beyond sentry
itself is a mistake

On Tue Dec 30 2014 at 10:17:48 AM Matt Robenolt [email protected]
wrote:

I think most of the use cases of CSP are valid. :)

The noisy/gross things are from browser plugins. Those are things I'd want
to ignore. But since they don't go through a client, we don't have an
opportunity to filter them out.

One of my ideas was to always relay through a raven-python or something
similar to do filter on _your_ server first, and raven-python could
expose a url to their Django app. But that's obviously very limited to
raven-python and Django.

—
Reply to this email directly or view it on GitHub
https://github.com/getsentry/sentry/issues/729#issuecomment-68381400.

[dcramer]

I think at some point we will have to build a proxy in front of Sentry (rather than relying on nginx hacks/etc) which can help solve this. That said, I agree with just pushing straight to the server. It's fairly fast to accept events already, but at some point we'll want to allow "block these kinds of events" which isn't cheap enough to do today.

[mattrobenolt]

it seems like expecting someone to have a server component beyond sentry itself is a mistake

I agree, which is why this is still open. :)

[dcramer]

FWIW I'm willing to support this immediately with the caveat that we dont suggest to the world they use it yet.

I dont have time to build anything yet, but I believe this is almost entirely "handle CSP reports explicitly at the store endpoint". That code is fairly old so it's not the most editable part of Sentry, but I dont think it would be extremely difficult.

The big question is "Do we need any custom interfaces for this data?" If we do then we should first spec those out, as I'm not a big fan of changing that side of Sentry so I want to do it with a solid grasp of whats required, and if it can be filled by other means.

[alex]

Do you mean interfaces as in user interfaces, or programmatic ones?

On Tue Dec 30 2014 at 10:27:12 AM David Cramer [email protected]
wrote:

FWIW I'm willing to support this immediately with the caveat that we dont
suggest to the world they use it yet.

I dont have time to build anything yet, but I believe this is almost
entirely "handle CSP reports explicitly at the store endpoint". That code
is fairly old so it's not the most editable part of Sentry, but I dont
think it would be extremely difficult.

The big question is "Do we need any custom interfaces for this data?" If
we do then we should first spec those out, as I'm not a big fan of changing
that side of Sentry so I want to do it with a solid grasp of whats
required, and if it can be filled by other means.

—
Reply to this email directly or view it on GitHub
https://github.com/getsentry/sentry/issues/729#issuecomment-68382246.

[dcramer]

@alex I mean the ones where we store the structured data (sentry.interfaces).

Mostly "if we capture this, how can we represent data in the most reusable fashion, ideally reusing existing stuff (but unlikely)"

[dcramer]

For posterity http://sentry.readthedocs.org/en/latest/developer/interfaces/index.html

[jvehent]

Did this feature ever get implemented?

[dcramer]

@jvehent it hasn't. We're still planning to explore this but historically we've heard lots of negatives on CSP from people using it (noisy and limited).

[oreoshake]

It's gotten much, much better over time.

[mattrobenolt]

I'm going to take a peek at this stuff over the weekend and see what this looks like in Sentry now as a proof of concept. Not sure if we'll commit to using it, but I haven't looked at CSP reporting since 2013. :)

[oreoshake]

:metal: well a lot has changed since 2013. Except for safari. Safari's implementation is awful and is probably closer to what you experienced globally in 2013. I highly recommend discarding safari CSP reports by default.

[mattrobenolt]

Leaving this here for reference: https://developer.mozilla.org/en-US/docs/Web/Security/CSP/Using_CSP_violation_reports

[oreoshake]

That isn't up to date, namely it doesn't reference the effective-directive that Chrome/Opera send and is part of the Level 2 spec sample violation report.

This list includes everything you should expect to see in the various reports. Additionally, Firefox sends an incredibly useful but non-standard value for script-sample. This can be used to detect plugins like lastpass (which will account for a large number of violations).

[mattrobenolt]

Awesome, thanks for the information!

[mattrobenolt]

For those that are interested in testing this, can you reach out to me? matt @ getsentry.com.

We're looking to get some community feedback before we unleash this into the wild. :)

[mattrobenolt]

Also, btw, thanks everyone for nudging this along!

[jquacinella]

Any thoughts on when this might be released? Could be willing to test

[dcramer]

@jquacinella its live for early adopters -- go to organization settings, and tick the 'Early Adopters' box. From there head to [Project] > Settings > CSP Reports (under Data)

[brianmhunt]

😎 Will Public Key Pin reports be handled appropriately by the CSP handler?

[mattrobenolt]

We do not support Public Key Pin violations yet. The demand is very very low. It'll likely happen at some point. See https://github.com/getsentry/sentry/issues/2165

[rohitpaulk]

So we're trying this out at Gratipay. Bit confused - where do we get to see reports in the Sentry UI? Are they supposed to be shown in the 'Issues' page?

[rohitpaulk]

^ False alarm, we're seeing a 403 forbidden when submitting CSP reports. Will update here if it doesn't work once we resolve that