Impact

When WebOb normalizes the HTTP Location header to include the request hostname, it does so by parsing the URL that the user is to be redirected to with Python's urllib.parse, and joining it to the base URL. urlsplit (called internally by urljoin) however treats a // at the start of a string as a URI without a scheme, and then treats the next part as the hostname. urljoin will then use that hostname from the second part as the hostname replacing the original one from the request.

In a previous advisory https://github.com/Pylons/webob/security/advisories/GHSA-mg3v-6m49-jhp3 an attempt to fix this was made by forcing the replacement of // with /%2f, however this did not take into account that since Python 3.10 urlsplit internally strips ASCII tab, carriage return, and newline characters from the string, so /\t/attacker.com gets turned into //attacker.com and the attacker is able to bypass the changes introduced in that previous advisory, thereby bringing back the problem that was attempted to be fixed.

>> parse.urlparse("//attacker.com/some/path") ParseResult(scheme='', netloc='attacker.com', path='/some/path', params='', query='', fragment='')

WebOb uses urljoin to take the request URI and join the redirect location to it, so assuming the request URI is https://example.org/ and the URL to redirect to is /\t/attacker.com/some/path/:

>> parse.urljoin("https://example.org/", "/\t/attacker.com/some/path/") 'https://attacker.com/some/path/'

Which redirects from example.org where we want the user to stay to attacker.com.

Patches

This issue has been fixed in WebOb 1.8.10.

Workarounds

Any use of the Response class that includes a location can be rewritten to make sure to always pass a full URI that includes the hostname to redirect the user to, or to validate that the redirect target starts with a scheme (e.g. http:// or https://) before assigning to Response.location.

References

- https://github.com/Pylons/webob/security/advisories/GHSA-mg3v-6m49-jhp3 - CVE-2024-42353

Thanks

- Caleb Brown of Google

Impact

When WebOb normalizes the HTTP Location header to include the request hostname, it does so by parsing the URL that the user is to be redirected to with Python's urlparse, and joining it to the base URL. urlparse however treats a // at the start of a string as a URI without a scheme, and then treats the next part as the hostname. urljoin will then use that hostname from the second part as the hostname replacing the original one from the request.

>> parse.urlparse("//example.com/test/path") ParseResult(scheme='', netloc='example.com', path='/test/path', params='', query='', fragment='')

WebOb uses urljoin to take the request URI and joining the redirect location, so assuming the request URI is: https://example.org//example.com/some/path, and the URL to redirect to (for example by adding a slash automatically) is //example.com/some/path/ that gets turned by urljoin into:

>> parse.urljoin("https://example.org//attacker.com/some/path", "//attacker.com/some/path/") 'https://attacker.com/some/path/'

Which redirects from example.org where we want the user to stay to attacker.com

Patches

This issue is patched in WebOb 1.8.8

Older versions of WebOb continue to be vulnerable to this issue, and should be avoided.

Workarounds

Any use of the Response class that includes a location can be rewritten to make sure to always pass a full URI that includes the hostname to redirect the user to.

Thanks

- Sara Gao

This issue was reported via the Pylons Project Security List