[Mediawiki-api] [Mediawiki-api-announce] Insecure (non-HTTPS) API Requests to become unsupported starting 2016-06-12

Brandon Black bblack at wikimedia.org
Fri May 13 22:34:20 UTC 2016

* All access to Wikimedia production sites/APIs should use https://
URLs, not http:// -- your bot/tool will break in the near future if it
does not!
* 2016-06-12 - insecure access is unsupported; starting on this date
we plan to break (deny with 403) 10% of all insecure requests randomly
as a wake-up call.
* 2016-07-12 - we plan to break all insecure requests.

Hi all,

As you may remember, all production Wikimedia wikis switched to
HTTPS-only for all canonical domainnames nearly a year ago:

Since way back then, we've been forcing insecure HTTP requests to our
canonical domains over to HTTPS by using redirects and
Strict-Transport-Security, which is effective for the vast majority of
access from humans using browsers and apps.

In the time since, we've been chasing down various corner-case issues
where loopholes may arise in our HTTPS standards and enforcement.  One
of the most-difficult loopholes to close has been the "Insecure POST"
loophole, which is discussed in our ticket system here:
https://phabricator.wikimedia.org/T105794 .

To briefly recap the "Insecure POST" issue:

* Most of our humans using browser UAs are not affected by it.  They
start out doing GET traffic to our sites, their GETs get redirected to
HTTPS if necessary, and then any POSTs issued by their browser use
protocol-relative URIs which are also HTTPS.
* However, many automated/code UAs (bots, tools, etc) access the APIs
using initial POST requests to hardcoded service URLs using HTTP
(rather than HTTPS).
* For all of the code/library UAs out there in the world, there is no
universally-compatible way to redirect them to HTTPS.  There are
different ways that work for some UAs, but many UAs used for APIs
don't handle redirects at all.
* Regardless of the above, even if we could reliably redirect POST
requests, that doesn't fix the security problem like it does with GET.
The private data has already been leaked in the initial insecure
request before we have a chance to redirect it.  If we did some kind
of redirect first, we'd still just be putting off the inevitable
future date where we have to go through a breaking transition to
secure the data.

Basically, we're left with no good way to upgrade these insecure
requests without breaking them.  The only way it gets fixed is if all
of our API clients in the world use explicit https:// URLs for
Wikimedia sites in all of their code and configuration, and the only
way we can really force them to do so is to break insecure POST
requests by returning a 403 error to tools that don't.

Back in July 2015, I began making some efforts to statistically sample
the User-Agent fields of clients doing "Insecure POST" and tracking
down the most-prominent offenders.  We were able to find and fix many
clients along the way since.

A few months ago Bryan Davis got us further when he committed a
MediaWiki core change to let our sites directly warn offending
clients.  I believe that went live on Jan 29th of this year (
https://gerrit.wikimedia.org/r/#/c/266958 ).  It allows insecure POSTs
to still succeed, but sends the clients a standard warning that says
"HTTP used when HTTPS was expected".

This actually broke some older clients that weren't prepared to handle
warnings at all, and caused several clients to upgrade.  We've been
logging offending UAs and accounts which trigger the warning via
EventLogging since then, but after the initial impact the rate
flattened out again; clients and/or users that didn't notice the
warning fairly quickly likely never will.

Many of the remaining UAs we see in logs are simply un-updated.  For
example, https://github.com/mwclient/mwclient switched to
HTTPS-by-default in 0.8.0, released in early January, but we're still
getting lots of insecure POST from older mwclient versions installed
out there in the world.  Even in cases where the code is up to date
and supports HTTPS properly, bot/tool configurations may still have
hardcoded http:// site config URLs.

We're basically out of "soft" ways to finish up this part of the HTTPS
transition, and we've stalled long enough on this.

** 2016-06-12 is the selected support cutoff date **

After this date, insecure HTTP POST requests to our sites are
officially unsupported.  This date is:

* A year to day after the public announcement that our sites are HTTPS only
* ~ 11 months after we began manually tracking down top offenders and
getting them fixed
* ~ 4 months after we began sending warning messages in the response
to all insecure POST requests to the MW APIs
* ~ 1 month after this email itself

On the support cutoff date, we’ll begin emitting a “403 Insecure POST
Forbidden - use HTTPS” failure for 10% of all insecure POST traffic
(randomly-selected).  Some clients will retry around this, and
hopefully the intermittent errors will raise awareness more-strongly
than the API warning message and this email did.

A month later (two months out from this email) on 2016-07-12 we plan
to break insecure access completely (all insecure requests get the 403

In the meantime, we'll be trying to track down offending bots/tools
from our logs and trying to contact owners who haven't seen these
announcements.  Our Community team will be helping us communicate this
message more-directly to affected Bot accounts as well.

Thank you all for your help during this transition!
-- Brandon Black
Sr Operations Engineer
Wikimedia Foundation

Mediawiki-api-announce mailing list
Mediawiki-api-announce at lists.wikimedia.org

More information about the Mediawiki-api mailing list