Partial password leak over DNS on HTTP redirect
Project curl Security Advisory, June 24th 2020 - Permalink
libcurl can be tricked to prepend a part of the password to the host name before it resolves it, potentially leaking the partial password over the network and to the DNS server(s).
libcurl can be given a username and password for HTTP authentication when
requesting an HTTP resource - used for HTTP Authentication such as Basic,
Digest, NTLM and similar. The credentials are set, either together with
CURLOPT_USERPWD or separately with
CURLOPT_PASSWORD. Important detail: these strings are given to libcurl as
plain C strings and they are not supposed to be URL encoded.
In addition, libcurl also allows the credentials to be set in the URL, using
the standard RFC 3986 format:
http://user:password@host/path. In this case,
the name and password are URL encoded as that's how they appear in URLs.
If the options are set, they override the credentials set in the URL.
Internally, this is handled by storing the credentials in the "URL object" so that there is only a single set of credentials stored associated with this single URL.
When libcurl handles a relative redirect (as opposed to an absolute URL redirect) for an HTTP transfer, the server is only sending a new path to the client and that path is applied on to the existing URL. That "applying" of the relative path on top of an absolute URL is done by libcurl first generating a full absolute URL out of all the components it has, then it applies the redirect and finally it deconstructs the URL again into its separate components.
This security vulnerability originates in the fact that curl did not correctly
URL encode the credential data when set using one of the
options described above. This made curl generate a badly formatted full URL
when it would do a redirect and the final re-parsing of the URL would then go
bad and wrongly consider a part of the password field to belong to the host
The wrong host name would then be used in a name resolve lookup, potentially leaking the host name + partial password in clear text over the network (if plain DNS was used) and in particular to the used DNS server(s).
The password leak is triggered if an at sign (
@) is used in the password
field, like this:
passw@rd123. If we also consider a user
dan, curl would
generate a full URL like:
... while a correct one should have been:
... when parsing the wrongly generated URL, libcurl would end up with user
dan and password
passw talking to the host
bad host name would then be passed on to the name resolver function in use
(and for all typical cases return a "cannot resolve host name" error).
There's no hint in the name resolve as to how large portion of the password
that is actually prepended to the host name (ie an observer won't know how
much data there was on the left side of the
@), but it can of course be a
significant enough clue for an attacker to figure out the rest.
We are not aware of any exploit of this flaw.
Requirements to trigger this flaw.
- a password set with a
- an HTTP transfer
- a relative redirect that curl follows (
This flaw can happen to users of the curl tool as well as for applications using libcurl.
This bug was reported and inadvertently fixed and pushed to the public source respository before anyone realized its security impact.
The effects of this flaw is somewhat reduced if DNS-over-HTTPS is used, since then at least the name won't be observable on the network by a passive by-stander but only by the DoH server.
The Common Vulnerabilities and Exposures (CVE) project has assigned the name CVE-2020-8169 to this issue.
CWE-200: Exposure of Sensitive Information to an Unauthorized Actor
Severity: 5.5 (Medium)
libcurl is used by many applications, but not always advertised as such.
We suggest you take one of the following actions immediately, in order of preference:
A - Upgrade curl to version 7.71.0
B - Apply the patch on your libcurl version and rebuild
C - Disable
CURLOPT_FOLLOWLOCATION or redirects to HTTP(S).
This issue was first reported to the curl project on May 14, 2020. The initial fix was done, verified and pushed to git on the same day. (As a regular non-security related fix.)
On May 15, 2020, the bug was reported again but then with the security impact highlighted.
This advisory was posted on June 24th 2020.
The security issue was reported by Marek Szlagor. The initial bug report was done by Gregory Jefferis and Jeroen Ooms. Patched by Daniel Stenberg.
Thanks a lot!