Table of Contents

1) Introduction

sslh is a protocol demultiplexer that allows to provide different types of services on the same network port. To achieve this, sslh performs heuristic analysis of the initial network data arriving on a connection, and forwards all further traffic to a matching service on the local system. A typical use case is to serve both SSL and SSH connections (hence the name) on port 443, to accommodate corporate firewall restrictions.

In April 2025 we conducted a review of sslh, mostly due to the fact that it processes all kinds of network protocols and is implemented in the C programming language, which is known to be prone to memory handling errors. For this review we looked into release v2.2.1 of sslh. Bugfixes for the issues described in this report can be found in release v2.2.4.

The next section provides an overview of the sslh implementation. Section 3) describes two security relevant Denial-of-Service issues we discovered during our review. Section 4) discusses some non-security relevant findings and remarks we gathered during our review. Section 5) looks into the general resilience of sslh against high network load attacks. Section 6) provides a summary of our assessment of sslh.

2) Overview of sslh

sslh implements so-called probes to determine the type of service when a new TCP or UDP session is initiated. These probes inspect the first few bytes of incoming data until a positive or a negative decision can be made. Once a specific service type has been determined, all following traffic will be forwarded to a dedicated service running on localhost, without interpreting further data. sslh will only probe for those protocols that are actively configured, no other probes will be invoked without need.

sslh supports three different I/O models for handling network input. The choice of what model to use is made at compile time, which is why there can exist multiple sslh binaries, one for each I/O flavor. The following models exist:

  • a fork model implemented in sslh-fork.c. In this model, a separate process is forked for each newly incoming TCP connection. The forked process obtains ownership of the TCP connection, handles related I/O, and exits when the connection ends. UDP protocols are not supported in this model.
  • a select model implemented in sslh-select.c. In this model, file descriptors are monitored in a single process using the select() system call. This model also supports UDP protocols: for this purpose, all data originating from the same source address are considered to be part of the same session. A dedicated socket is created for each new session sslh detects.
  • an implementation based on libev implemented in sslh-ev.c. This variant outsources the I/O management details to the third party library. This also supports UDP protocols in a similar way to the select model described earlier.

The different probes implemented in sslh were one of the focus areas during our review. sslh runs with lowered privileges and systemd hardenings enabled, thus privilege escalation attack vectors will only have limited impact. An area that is still important in spite of these protections is Denial-of-Service, which we looked into as well.

3) Security Issues

3.1) File Descriptor Exhaustion Triggers Segmentation Fault (CVE-2025-46807)

As part of our investigation of Denial-of-Service attack vectors, we looked into what happens when a lot of connections are created towards sslh and, as a result, file descriptors are exhausted. While the sslh-fork variant manages file descriptor exhaustion quite well, the other two variants have issues in this area. This especially affects UDP connections that need to be tracked on application level, since there is no concept of a connection on protocol level.

For each connection, sslh maintains a timeout after which the connection is terminated if the type of service could not be determined. The sslh-select implementation only checks UDP timeouts when there is network activity, otherwise the file descriptors that are created for each UDP session stay open. Due to this, an attacker can create enough sessions to exhaust the 1024 file descriptors supported by default by sslh, thereby making it impossible for genuine clients to connect anymore.

Even worse, when the file descriptor limit is encountered, sslh crashes with a segmentation fault, as it attempts to dereference new_cnx, which is a NULL pointer in this case. Therefore, this issue represents a simple remote Denial-of-Service attack vector. The segmentation fault also happens when the admin configures the udp_max_connections setting (or command line switch), as the NULL pointer dereference is reached in this context as well.

To reproduce this, we tested the openvpn probe configured for UDP. On the client side we created many connections where each connection only sends a single 0x08 byte.

We did not check the sslh-ev implementation very thoroughly, because it depends on the third party libev library. The behaviour is similar to the sshl-select variant, though. UDP sockets are seemingly never closed again.

Bugfix

Upstream fixed this issue in commit ff8206f7c which is part of the v2.2.4 release. While the segmentation fault is fixed with this change, UDP sockets potentially still stay open for a longer time until further traffic is processed by sslh, which triggers the socket timeout logic.

3.2) Misaligned Memory Accesses in OpenVPN Protocol Probe (CVE-2025-46806)

In the UDP code path of is_openvpn_protocol(), if clauses like this can be found:

    if (ntohl(*(uint32_t*)(p + OVPN_HARD_RESET_PACKET_ID_OFFSET(OVPN_HMAC_128))) <= 5u)

This dereferences a uint32_t* that points to memory located 25 bytes after the start of the heap allocated network buffer. On CPU architectures like ARM this will cause a SIGBUS error, and thus represents a remote DoS attack vector.

We reproduced this issue on a x86_64 machine by compiling sslh with
-fsanitize=alignment. By sending a sequence of at least 29 0x08 bytes, the following diagnostic is triggered:

probe.c:179:13: runtime error: load of misaligned address 0x7ffef1a5a499 for type 'uint32_t', which requires 4 byte alignment
0x7ffef1a5a499: note: pointer points here
 08 08 08  08 08 08 08 08 08 08 08  08 08 08 08 08 08 08 08  08 08 08 08 08 08 08 08  08 08 08 08 08
              ^
probe.c:185:13: runtime error: load of misaligned address 0x7ffef1a5a49d for type 'uint32_t', which requires 4 byte alignment
0x7ffef1a5a49d: note: pointer points here
 08 08 08 08 08 08 08  08 08 08 08 08 08 08 08  08 08 08 08 08 08 08 08  08 08 08 08 08 08 08 08  08

Bugfix

The usual fix for this problem in protocol parsing is to memcpy() the integer data into a local stack variable instead of dereferencing the pointer into the raw network data. This is what upstream did in commit 204305a88fb3 which is part of the v2.2.4 release.

4) Other Findings and Remarks

4.1) Missing Consideration of Short Reads on TCP Streams

A couple of probes don’t consider short reads when dealing with the TCP protocol. For example in is_openvpn_protocol() the following code is found in the TCP code path:

    if (len < 2)
        return PROBE_AGAIN;

    packet_len = ntohs(*(uint16_t*)p);
    return packet_len == len - 2;

If less than two bytes have been received, then the function indicates PROBE_AGAIN, which is fine. After the supposed message length has been parsed into packet_len, the probe only succeeds if the complete message has been received by now, otherwise the function returns 0 which equals PROBE_NEXT.

Similar situations are found in is_teamspeak_protocol() and is_msrdp_protocol(). While it may be unlikely that such short reads occur often with TCP, it is still formally incorrect and could lead to false negative protocol detection in a number of cases.

Bugfix

Based on experience upstream believes that this is not an issue in practice currently, since no bug reports in this area have appeared. For this reason this is not a priority for upstream at the moment.

4.2) Likelihood of False Positive Probe Results

A couple of probe functions rely on very little protocol data to come to a positive decision. For example is_tinc_protocol() indicates a match if the packet starts with the string
" 0". In is_openvpn_protocol() any packet that stores the packet length in the first two bytes in network byte order is considered a match, which is probably the case for quite a few network protocols.

Security-wise this is not relevant, because the services these packets are being forwarded to have to be able to deal with whatever data is sent to them, even if it is destined for a different type of service. From a perspective of correct probe implementation it could lead to unexpected behaviour in some situations, however (especially when a lot of protocols are multiplexed over the same sslh port). We suggested upstream to try and base probe decisions on more reliable heuristics to avoid false positives.

Bugfix

Similar to section 4.1) upstream does not believe that this is a major issue for users at the moment, hence there are no immediate changes to the code base to address this.

4.3) Parsing of Potentially Undefined Data in is_syslog_protocol()

The following code is found in is_syslog_protocol():

    res = sscanf(p, "<%d>", &i);
    if (res == 1) return 1;

    res = sscanf(p, "%d <%d>", &i, &j);
    if (res == 2) return 1;

The sscanf() function does not know about the boundaries of the incoming network data here. Very short reads like a 1 byte input will cause sscanf() to operate on undefined data, found in the buffer allocated on the heap in defer_write().

Bugfix

For a quick bugfix we suggested to explicitly zero terminate the buffer by allocating an extra byte after the end of the payload. Running sscanf() to parse integers found in untrusted data could be considered a bit on the dangerous side, however, thus we suggested to generally try and change this into more careful code.

Upstream fixed this in commit ad1f5d68e96, which is part of the v2.2.4 release. The bugfix is along the lines of our suggestion and also adds an additional sanity check for the integer which is parsed from the network data.

5) Resilience of sslh Against High Network Load Attacks

A general-purpose network service like sslh can be sensitive to resource depletion attacks, such as the aforementioned file descriptor exhaustion issue. The sslh-fork implementation spawns a new process for each incoming TCP connection, which brings to mind the possibility to consume excessive resources not only on a per-process scope, but also on a system-wide scope. By creating a large amount of connections towards sslh, a “fork bomb” effect could be achieved. When a “fork bomb” is executed locally on Linux then this often still causes an inaccessible system even today, when there are no strict resource limits in place. Achieving something like this remotely would be a major DoS attack vector.

sslh-fork implements a timeout for each connection, which is based on the select() system call. If the probing phase does not come to a decision before the timeout occurs, then the connection is closed again. By default this timeout is set to five seconds. Since sslh-fork creates a new process for each newly incoming connection, there is no limit of 1024 file descriptors being opened by sslh. In theory an attacker could attempt to exceed the system-wide file descriptor and/or process limit by creating an excessive amount of connections.

The default timeout enforcement of five seconds means that the attack is quite limited, however. During our tests we were not able to create more than about 5,000 concurrent sslh-fork processes. This creates quite a bit of system load, but does not expose any critical system behaviour on an average machine.

Even though the current situation is acceptable, it could be considered to offer an application level limit for the amount of parallel connections. For UDP there exists a udp_max_connections setting already, but not for TCP.

Bugfix

In discussions with upstream it was agreed that proper protection from such Denial-of-Service attacks is best achieved on the end of the administrator, who can for example configure Linux cgroup constraints. Upstream is still considering to add a tcp_max_connections setting to limit the maximum amount of parallel TCP connections in the future.

6) Summary

Overall we believe sslh is in good shape. There is little attack surface, and hardenings are in place by default. With the two remote DoS vectors 3.1) and 3.2) fixed, it should be safe to use sslh in production. Users that are worried about more complex DoS attacks should additionally consider customizing their setup to enforce resource consumption limits on operating system level.

There is some danger of false positive or false negative probe outcomes as outlined in sections 4.1) and 4.2). These seem not to have occurred a lot in practice yet, and it is a trade-off towards simplicity and efficiency in the current implementation of sslh.

7) Timeline

2025-04-25 We privately reported the findings to the author of sslh by email, offering coordinated disclosure.
2025-05-06 We discussed details about the reported issues and possible CVE assignments. The issues were kept private for the time being.
2025-05-08 We assigned two CVEs from our pool for the issues and shared them with upstream.
2025-05-25 The upstream author informed us about bugfixes that have already been published in the sslh GitHub repository and about an upcoming release containing the fixes.
2025-05-28 Release v2.2.4 containing the fixes was published.
2025-06-13 We published this report.

8) References