Yubico libyubihsm Vulnerabilities (CVE-2021-27217, CVE-2021-32489)
ID: CVE-2021-27217, CVE-2021-32489
I’ve continued my research on weaknesses in the Yubico HSM library and recently found additional vulnerabilities in it through fuzzing.
A malicious HSM device with authentication can trigger out of bounds read operations (CVE-2021-27217) on the host that may lead to segmentation faults and crash the program.
This article will describe the issues and give some background information on how they were found.
I’m a freelance Security Consultant and currently available for new projects. If you are looking for assistance to secure your projects or organization, contact me.
Please see the previous yubihsm-shell article for a summary of the automated testing approach and technical background.
Notably, the necessary fuzzer progress for the new discovery was achieved by weakening the normal message authentication checks in the code sufficiently to make the fuzzer pass them with some inputs. This allowed the fuzzer + sanitizer combination to discover problematic code located behind that barrier.
Without this strategic adjustment, passing the authentication checks with fuzzer-provided input would otherwise be equal to a brute-force attack on the valid message authentication code (a specific 64 bit value), which is very hard for a generic fuzzer to do.
Note that this step also corresponds to a change in the attacker model from “man-in-the-middle without knowledge of secrets” to “malicious HSM device with knowledge of secrets”, since the latter is required to accurately compute the required MAC. More on this in later sections.
Conceptual overview of the fuzzing setup:
yubishm- +-----------------+ shell | libyubihsm | fuzzer | | | USB handling | +--------^--------+ | | +--------+--------+ | mod. USB backend| | libfuzzer | +-----------------+
The underlying code issue is near the location of CVE-2020-24388 and triggered similarly, so I recommend reading the original issue description to understand this issue.
In short, the libyubihsm code and the HSM are interacting via USB messages, which are commonly transported over other network protocols in remote setups via the yubihsm-connector.
After sending a secure message to the HSM via
_send_secure_msg(), the host program expects the HSM to respond with an authenticated and encrypted message that contains some response data related to the request, for example a status code and response data for the requested HSM action.
The response message is initially checked as follows:
For the previous CVE-2020-24388 issue, messages with
response_msg.st.len < SCP_MAC_LEN - 3 caused dangerous
memcpy() problems during the early parsing steps.
In particular, the problematic
memcpy() happened before the message authentication field was checked, so an unauthenticated man-in-the-middle attacker could trigger it.
This problem was fixed in libyubihsm 2.0.3 by - among other things - requiring
response_msg.st.len >= SCP_MAC_LEN - 3 as seen above.
Overall, the steps towards the MAC comparison now appear to be safe from this kind of attack based on problematic length fields.
However, it turns out that the 2.0.3 patches failed to prevent length-related issues that happen after the MAC comparison. Let’s assume that libyubihsm is interacting with a malicious HSM device that has access to the correct secrets for authentication & encryption measures between the HSM and host.
The program looks as follows after the authentication stage:
Since the libyubihsm <> HSM interaction happens in several logical message flows called sessions, the code first checks if the response message is actually meant for the current session to prevent mixups between concurrent sessions due to race conditions. The main message payload within the data section of the message is then decrypted and any remaining AES padding is removed.
aes_remove_padding() function that is called at the end of the code excerpt makes dangerous assumptions about the memory contents and length variable that it operates on:
As you can see, the function mainly cares about the memory contents at
in, and not at all about the boundary conditions related to
If the function is called with
len = 0 or a pointer to a buffer filled with
0x00 bytes, then it will continue reading data and walking over memory in front of the provided buffer until it encounters a non-null data byte. This type of out-of-bounds memory access will not leak a lot of information, but ignores any memory boundaries that are in place along the way.
Depending on the memory contents in front of the buffer and general memory layout, this can cause the libyubihsm program to be terminated with a segmentation fault.
Due to the way that the message parsing logic is written, it is in fact possible for a crafted message to reach
aes_remove_padding() with an
out_len = 0 value.
response_msg.st.len = 9 are accepted and
SCP_MAC_LEN + 1 = 9 is subtracted from that length value during the parsing steps.
A problematic message would therefore consist of
- an acceptable response code such as
- the length value
- the correct session ID
- a valid 8-byte MAC code
During parsing, the code will then call
aes_remove_padding() on a 0-byte inner data payload segment with length 0, which triggers the out-of-bounds read operations.
Since the attack scenario assumes that the HSM is malicious and also knows the encryption keys, it is also possible for the HSM to send a longer message of size
response_msg.st.len > 9 where the encrypted inner data payload section is chosen so that it decodes to all
0x00 bytes. This causes the
aes_remove_padding() call to step over the provided payload buffer and then perform the same out of bounds reads as with the previous case.
This means that both additional length checks as well as a better
aes_remove_padding() have to be put in place to avoid this.
During execution, there are three possible outcomes:
Depending on compiler settings and other factors, the program memory can be such that
aes_remove_padding()walks until reaching external memory boundaries of the program, causing a segmentation fault. This is the problematic outcome that represents a denial of service impact.
If the conditions are not right for a segmentation fault (the attacker has no influence over this), the program execution continues as
aes_remove_padding()returns with some
out_lenvalue. Due to the unsigned integer overflow, this is usually a large value which is exceeding the allocated buffer size. In this case, a defense condition that was meant for another fault condition is triggered. The function then returns with an explicit
YHR_BUFFER_TOO_SMALLerror, halting further processing and preventing any practical impact:
3. Like 2), but with
0 < out_len < *response_len. This has not been observed in practice and was not investigated further, but may cause additional memory problems.
While documenting CVE-2021-27217, I noticed a curious edge case.
response_msg.st.len = 9 looks like the shortest message length that can ever be accepted.
After all, the library code wants 1 byte of valid channel ID and 8 bytes for the MAC data, so shorter messages are rejected automatically, right?
Actually, no. Looking closely at the code showed that the two fields are read from different sides of the payload:
sid is read from the lowest memory position of the buffer:
MAC is read from the highest memory position of the buffer:
The two fields may actually overlap, meaning that the first MAC byte can be a valid session ID, making the message valid!
Let’s assume for a minute that both checks do pass on a message that was received with
response_msg.st.len = 8.
In this case, the
uint16_t out_len variable is decremented too far by the combination of
out_len -= 8 and
out_len -= 1.
As a result, it overflows to
This becomes a significant problem a few steps further into the parsing when running the AES decryption step:
aes_cbc_decrypt(), OpenSSL is now told to work on a buffer of size 65535 while the buffer handed to it is actually far smaller.
Unsurprisingly, this leads to a segmentation fault in OpenSSL, here shown in
Unlike the distinct
aes_remove_padding() based problem, this code path will always crash and does not depend on certain memory layouts.
This leaves us with the question: how can an attacker pass the session ID and message authentication checks in order to reach this case?
The slow and “naive” attack variant consists of the HSM behaving “harmlessly” by generally responding to request messages with non-malicious response messages. At the same time, it silently computes the potential MAC of a short attack message based on the current cryptographic parameters of the communication. Only if the first byte of this MAC matches the session ID of the request, it launches the attack and sends a short reply message.
Due to the cryptographic properties of the MAC, the relevant first MAC byte is basically changing in a random fashion and the average chance of a match with the current
sid value is 1/256 for each request.
So on average there is an attack opportunity every few hundred messages. This is plausible to reach in an automated setup where the HSM is queried regularly over some time.
However, this situation does put some limits on the attack for an HSM that is queried rarely.
After thinking about this problem for a bit, I figured out an optimized attack variant that does this with significantly less message request opportunities.
Our crafted message looks somewhat like this when plotted with the individual fields and offsets:
1 2 3 4 5 6 7 8 9 10 11 | byte position [cmd] | [length field] | field usage [ sid ] | [ MAC field ] |
Normally, the MAC is computed over
[cmd] [length field] [payload data without MAC section], but since the remaining payload data has shrunk to zero and the length is fixed in our attack,
the MAC is effectively just computed over the meta section
This appears somewhat hopeless at first since there apparently is no room for variations of the message, until one looks at the limits imposed on
cmd during the parsing:
We can change
cmd to any value except the error code
YHC_ERROR = 0x7f when trying to trigger the bug!
Since message authentication codes are supposed to change wildly for each variation of the message,
trying out the 255 different accepted variants for
cmd results in the excellent average chance of 255/256 =
~99,6% for a valid
sid <> first MAC byte match that allows the attack.
At worst, the malicious HSM has to wait out a few requests for an attack opportunity, but not a few hundred.
That’s a decent improvement.
Attack Scenario and Security Implications
Please note that the vulnerabilities are in libyubihsm on the host side. They can only be triggered by a trusted but malicious HSM or a man-in-the-middle attacker with similar knowledge of secrets. This is a very specific attack subclass that requires a lot of preparation and sophistication by the attacker. However, given that supply chain attacks are harder to defend against than most people realize (see enclosure and tamper evident seal issues of the YubiHSM2 packaging), it is beneficial to research and improve the defenses in this scenario as well.
If you are confident that your physical HSM2 modules are genuine or that your software stacks exclude libyubihsm then the described issues do not affect you.
Yubico has rated the issues as follows:
|ID||CVSS 3.1 Score||Parameters|
Note that the initial YSA-2021-01 advisory included inaccurate CVSS score information, but has been updated.
Proof Of Concept
The following patches simulate the attack with limited code changes to a vulnerable library version. This can be tested with an expendable HSM2 device to validate the error behavior without a complex hardware setup.
WARNING: use the following code at your own risk. Although not intended, please assume that this will PERMANENTLY overwrite data on the HSM device that you have connected.
- Rebuild libyubihsm and all relevant programs with the specified patch, example: vulnerable repo version
- Insert a test HSM device in factory settings
- Run a basic command to test the impact
- Variant A: yubihsm-connector is running on localhost, run
./yubihsm-shell -a blink-device -p password
- Variant B: direct USB connection, run
./yubihsm-shell -C yhusb:// -a blink-device -p password
If you observe connection problems, please make sure that your setup is working with the regular unmodified CLI.
As described previously, whether or not this leads to a segmentation fault is affected by compiler settings such as
-O0 and not expected to happen on all systems and configurations.
Second Issue Variant
./yubihsm-shell --connector=yhusb:// -a blink-device -p password Session keepalive set up to run every 15 seconds Created session 1 Segmentation fault
Due to previous disclosures on this product with Yubico, there was a lot of familiar ground, although the communication was not as fluid as with the last disclosures.
Time-wise, the patch response time remains close to the 90 day mark. This is likely connected to the YubiHSM2 SDK release process that focuses on major releases. Although faster patch times would be preferable to me as a researcher, in this particular case the vulnerabilities are not really grave enough as to require an urgent release.
Overall, the disclosure process was okay, but there is some room for improvement.
Relevant yubihsm-shell Sources
|Yubico upstream||GitHub||version 2.1.0, bundled in SDK release 2021.03||YSA-2021-01|
|2020-12-14||Disclosure of issue #1 to Yubico|
|2020-12-15||Yubico acknowledges receiving the report|
|2020-12-15||POC for issue #1 provided to Yubico|
|2021-01-04||Yubico confirms issue #1|
|2021-01-18||Yubico shares proposed fix|
|2021-01-18||Positive feedback on fix, recommendation of additional parsing changes|
|2021-02-05||Yubico communicates planned disclosure date: 2021-02-22|
|2021-02-13||Disclosure of issue #2 to Yubico|
|2021-02-16||Yubico communicates CVE-2021-27217 as ID for the primary issue|
|2021-02-19||Yubico postpones planned disclosure date|
|2021-02-25||Yubico communicates planned disclosure date: 2021-03-04|
|2021-03-04||Release of patched HSM2 SDK version and Yubico advisory YSA-2021-01|
|2021-03-04||Publication of this blog post|
|2021-03-10||Yubico updates YSA-2021-01|
|2021-05-10||Yubico obtains second CVE|
Note that the initial Yubico YSA-2021-01 advisory contained an incorrect publication date.
Yubico provided hardware as a bug bounty for this issue.