Copy of an email received on the Python Security Response team, 9 days ago. I consider that it's not worth it to have an embargo on this vulnerability, so I make it public.

Hi there,

I believe I've found a denial-of-service (DoS) bug in
urllib.request.AbstractBasicAuthHandler. To start, I'm operating on some
background information from this document: HTTP authentication
<https://developer.mozilla.org/en-US/docs/Web/HTTP/Authentication\>. The bug
itself is a ReDoS <https://www.regular-expressions.info/redos.html\> bug
causing catastrophic backtracking. To reproduce the issue we can use the
following code:

from urllib.request import AbstractBasicAuthHandler
auth_handler = AbstractBasicAuthHandler()
auth_handler.http_error_auth_reqed(
    'www-authenticate',
    'unused',
    'unused',
    {
        'www-authenticate': 'Basic ' + ',' * 64 + ' ' + 'foo' + ' ' +
'realm'
    }
)

The issue itself is in the following regular expression:

rx = re.compile('(?:.*,)*[ \t]*([^ \t]+)[ \t]+'
                'realm=(["\']?)([^"\']*)\\2', re.I)

In particular, the (?:.*,)* portion. Since "." and "," overlap and there
are nested quantifiers we can cause catastrophic backtracking by repeating
a comma. Note that since AbstractBasicAuthHandler is vulnerable, then both
HTTPBasicAuthHandler and ProxyBasicAuthHandler are as well because they
call http_error_auth_reqed. Building from the HTTP authentication document
above, this means a server can send a specially crafted header along with
an HTTP 401 or HTTP 407 and cause a DoS on the client.

I won't speculate on the severity of the issue too much - you will surely
understand the impact better than I will. Although, the fact that this is
client-side as opposed to server-side appears to reduce the severity,
however the fact that it's a security-sensitive context (HTTP
authentication) may raise the severity.

One possible fix would be changing the rx expression to the following:

rx = re.compile('(?:[^,]*,)*[ \t]*([^ \t]+)[ \t]+'
                'realm=(["\']?)([^"\']*)\\2', re.I)

This removes the character overlap in the nested quantifier and thus
negates the catastrophic backtracking.

Let me know if you have any questions or what the next steps are from here.
Thanks for supporting Python security!

--
Matt Schwager

I added this vulnerability to the following page to track fixes in all Python supported branches:
https://python-security.readthedocs.io/vuln/urllib-basic-auth-regex.html

CVE-2020-8492 has been assigned to this vulnerability:
https://cve.mitre.org/cgi-bin/cvename.cgi?name=2020-8492

Isn't this a duplicate of bpo-38826 ?

Isn't this a duplicate of bpo-38826 ?

Oh right. I marked it as a duplicate of this issue.

bench_parser.py: Benchmark for AbstractBasicAuthHandler.http_error_auth_reqed().

Instead of

repeat_10_3 = 'Basic ' + ', ' * (10 ** 3) + simple

in the benchmark, try

repeat_10_3 = 'Basic ' + ', ' * (10 ** 3) + 'A'

Ooooh, I see. I didn't measure the performance of the right header. I re-run a benchmark using the HTTP header (repeat=15):

    header = 'Basic ' + ', ' * 15 + 'A'

Now I see a major performance difference. Comparison between master ("ref") and PR 18284 ("fix"):

Mean +- std dev: [ref] 88.9 ms +- 2.4 ms -> [fix] 17.5 us +- 0.7 us: 5083.23x faster (-100%)

So the worst case is now way faster: more than 5000x faster!

It's even possible to go up to repeat=10**6 characters, it still takes less than 1 seconds: 412 ms +- 19 ms.

On the master branch, repeat=20 already takes around 3 seconds... The slowdown is exponential with repeat increase.

New changeset 0b297d4 by Victor Stinner in branch 'master':
bpo-39503: CVE-2020-8492: Fix AbstractBasicAuthHandler (GH-18284)
0b297d4

New changeset ea9e240 by Miss Islington (bot) in branch '3.8':
bpo-39503: CVE-2020-8492: Fix AbstractBasicAuthHandler (GH-18284) (GH-19296)
ea9e240

New changeset b57a736 by Miss Islington (bot) in branch '3.7':
bpo-39503: CVE-2020-8492: Fix AbstractBasicAuthHandler (GH-18284) (GH-19297)
b57a736

New changeset 69cdeeb by Victor Stinner in branch '3.6':
bpo-39503: CVE-2020-8492: Fix AbstractBasicAuthHandler (GH-18284) (GH-19304)
69cdeeb

New changeset 37fe316 by Victor Stinner in branch '3.5':
bpo-39503: CVE-2020-8492: Fix AbstractBasicAuthHandler (GH-18284) (bpo-19305)
37fe316

"headers" is a dict object? If so, the dict object does not seem to have no attribute "get_all".

No, it's not a dict object.

At the beginning of the issue, there is the following reproduction code:
from urllib.request import AbstractBasicAuthHandler
auth_handler = AbstractBasicAuthHandler()
auth_handler.http_error_auth_reqed(
'www-authenticate',
'unused',
'unused',
{
'www-authenticate': 'Basic ' + ',' * 64 + ' ' + 'foo' + ' ' +
'realm'
}
)

Here's the headers:

{
'www-authenticate': 'Basic ' + ',' * 64 + ' ' + 'foo' + ' ' +
'realm'
}

I think this is a dict object, so the current problem is fixed and no longer compatible with the previous usage?

This issue was a security vulnerability. It's now closed, please don't comment closed issues. If you consider that there is a regression, please open a new issue.