Recently, we found a critical vulnerability in StartCom’s new StartEncrypt tool, that allows an attacker to gain valid SSL certificates for domains he does not control. While there are some restrictions on what domains the attack can be applied to, domains where the attack will work include google.com, facebook.com, live.com, dropbox.com and others.
StartCom, known for its CA service under the name of StartSSL, has recently released the StartEncrypt tool. Modeled after LetsEncrypt, this service allows for the easy and free installation of SSL certificates on servers. In the current age of surveillance and cybercrime, this is a great step forwards, since it enables website owners to provide their visitors with better security at small effort and no cost.
However, there is a lot that can go wrong with the automated issuance of SSL certificates. Before someone is issued a certificate for their domain, say computest.nl, the CA needs to check that the request is done by someone who is actually in control of the domain. For “Extended Validation” certificates this involves a lot of paperwork and manual checking, but for simple, so-called “Domain Validated” certificates, often an automated check is used by sending an email to the domain or asking the user to upload a file. The CA has a lot of freedom in how the check is performed, but ultimately, the requester is provided with a certificate that provides the same security no matter which CA issued it.
So, StartEncrypt. In order to make the issuance of certificates easy, this tool runs on your server (Windows or Linux), detects your webserver configuration, and requests DV certificates for the domains that were found in your config. Then, the StartCom API does a HTTP request to the website at the domain you requested a certificate for, and checks for the presence of a piece of proof that you have access to that website. If the proof is found, the API returns a certificate to the client, which then installs it in your config.
However, it appears that the StartEncrypt tool did not receive proper attention from security-minded people in the design and implementation phases. While the client contains numerous vulnerabilities, one in particular allows the attacker to “trick” the validation step.
The first bug
The API checks if a user is authorized to obtain a certificate for a domain by downloading a signature from that domain, by default from the path
/signfile. However, the client chooses that path during a HTTP POST request to that API.
A malicious client can specify a path to any file on the server for which a certificate is requested. This means that, for example, anyone can obtain a certificate for sites like dropbox.com and github.com where users can upload their own files.
That’s not all
While this is serious, most websites don’t allow you to upload a file and then have it presented back to you in raw format like github and dropbox do. Another issue in the API however allows for much wider exploitation of this issue: the API follows redirects. So, if the URL specified in the “verifyRes” parameter returns a redirect to another URL, the API will follow that until it gets to the proof. Even if the redirect goes off-domain. Since the first redirect was to the domain that is being verified, the API considers the proof correct even if it is redirected to a different website.
This means that an attacker can obtain an SSL certificate for any website that either:
- Allows users to upload files and serves them back raw, or
- Has an “open redirect” vulnerafeature in it
Open redirects are pages that take a parameter that contains a URL, and redirect the user to it. This seems like a strange feature, but this mechanism is often implemented for instance in logout or login pages. After a successful logout, you will be redirected to some other page. That other page URL is included as a parameter to the logout page. For instance,
http://mywebsite/logout?returnURL=/login might redirect you to
/login after logout.
While many see open redirects as a vulnerability that needs fixing, others do not. Google for instance has excluded open redirects from their vulnerability reward program. By itself an open redirect is not exploitable, so it is understandable that many do not see it as a security issue. However, when combined with a vulnerability like the one present in StartEncrypt, the open redirect suddenly has great impact.
It’s actually even worse: the OAuth 2.0 specification practically mandates that an open redirect must be present in each implementation of the spec. For this reason, login.live.com and graph.facebook.com for instance contain open redirects.
When combining the path-bug with the open redirect, suddenly many more certificates can be obtained, like for google.com, paypal.com, linkedin.com, login.live.com and all those other websites with open redirects. While not every website has an open redirect feature, many do at some point in time.
That’s still not all
Apart from the vulnerability described above, we found some other vulnerabilities in the client while doing just a cursory check. We are only publishing those that according to StartCom have been fixed or are no issue. These include:
- The client doesn’t check the server’s certificate for validity when connecting to the API, which is pretty ironic for an SSL tool.
- The API doesn’t verify the content-type of the file it downloads for verification, so attackers can obtain certificates for any website where users can upload their own avatars (in combination with the above vulnerabilities of course)
- The API only involves the server obtaining the raw RSA signature of the nonce. The signature file doesn’t identify the key nor the nonce that was used. This opens up the possibility of a “Duplicate-Signature Key Selection” attack, just like what Andrew Ayer discovered in the ACME protocol while LetsEncrypt was in beta, see also this post. As long as the signfile is on a domain, an attacker can obtain the signfile, start a new authorization and obtain a nonce and then pick their RSA private key in such a way that their private key verifies their nonce.
- The private key on the server (
/usr/local/StartEncrypt/conf/cert/tokenpri.key) is saved with mode
0666, so world-readable, which means any local user can read or modify it.
All in all, it doesn’t seem like a lot of attention was paid to security in the design and implementation of this piece of software.
As a security company, we constantly do security research. Usually for paying customers. In this case however, we started looking at StartEncrypt out of interest and because of the high impact any issues have for the internet as a whole. That is also why we are disclosing this issue: we believe that CA’s need to become much more aware of the vital role they play in internet security and need to start taking their software security more serious.
Of course, we disclosed the issue to StartCom prior to publishing. The vulnerabilities we described were found in the Linux x86_64 binary version 188.8.131.52. The timeline:
June 22, 2016: issue discovered
June 23, 2016: issue disclosed to StartCom after obtaining email address by phone
June 23, 2016: StartCom takes API offline
June 28, 2016: StartCom takes API online again, incompatible with current client
June 28, 2016: StartCom updates the Windows-client on the download page
June 29, 2016: StartCom updates the Linux-client on the download page, keeping the version number at 184.108.40.206
June 30, 2016: StartCom informs Computest of which issues have been fixed
StartCom launched a tool that makes it easier to secure communications on the internet, which is something we applaud. In doing so however, they seem to have taken some shortcuts in engineering. Using their tool, an attacker is able to obtain certificates for other domains like google.com, linkedin.com, login.live.com and dropbox.com.
In our opinion, StartCom made a mistake by publishing StartEncrypt the way it is. Although they appreciated the issues for the impact they had and were swift in their response, it is apparent that too little attention was paid to security both in design (allowing the user to specify the path) and implementation (for instance in following redirects, static linking against a vulnerable library, and so on). Furthermore, they didn’t learn from the issues LetsEncrypt faced when in beta.
But the issue is broader. As users of the internet, we trust the CA’s to provide us with a base for trust upon which we do business and share our lives online. When a single CA publishes software with this level of security, the trust in the CA system as a whole is undermined.