# Hardening The Erlang/OTP SSH application is intended to be used in other applications as a library. Ensure the Erlang VM runs as a non-root OS user. All SSH services (shell, exec, SFTP) inherit the OS-level rights of the VM process. See [Terminology](terminology.md) for details on the rights model. Different applications using this library may have very different requirements. One application could be running on a high performance server, while another is running on a small device with very limited cpu capacity. For example, the first one may accept many users simultaneously logged in, while the second one wants to limit them to only one. That simple example shows that it is impossible to deliver the SSH application with default values on hardening options as well on other options that suit every need. The purpose of this guide is to discuss the different hardening options available, as a guide to the reader. Configuration in general is described in the [Configuration in SSH](configurations.md) chapter. ## DoS Resilience (Server) [](){: #resilience-to-dos-attacks } The following applies to daemons (servers). DoS (Denial of Service) attacks are hard to fight at the node level. Here are firewalls and other means needed, but that is out of scope for this guide. However, some measures could be taken in the configuration of the SSH server to increase the resilence. The options to use are: ### Counters and Parallelism - **[max_sessions](`m:ssh#hardening_daemon_options-max_sessions`)** - The maximum number of simultaneous sessions that are accepted at any time for this daemon. This includes sessions that are being authenticated. The default is that an unlimited number of simultaneous sessions are allowed. It is a good candidate to set if the capacity of the server is low or a capacity margin is needed. - **[max_channels](`m:ssh#hardening_daemon_options-max_channels`)** - The maximum number of channels that are accepted for each connection. The default is unlimited. - **[parallel_login](`m:ssh#hardening_daemon_options-parallel_login`)** - If set to false (the default value), only one login is handled at a time. If set to true, the number of simultaneous login attempts are limited by the value of the [max_sessions](`m:ssh#hardening_daemon_options-max_sessions`) option. ### Timeouts - **[hello_timeout](`t:ssh:hello_timeout_daemon_option/0`)** - If the client fails to send the first ssh message after a tcp connection setup within this time (in milliseconds), the connection is closed. The default value is 30 seconds. This is actually a generous time, so it can lowered to make the daemon less prone to DoS attacks. - **[negotiation_timeout](`t:ssh:negotiation_timeout_daemon_option/0`)** - Maximum time in milliseconds for the authentication negotiation counted from the TCP connection establishment. If the client fails to log in within this time the connection is closed. The default value is 2 minutes. It is quite a long time, but can lowered if the client is supposed to be fast like if it is a program logging in. - **[idle_time](`t:ssh:max_idle_time_common_option/0`)** - Sets a time-out on a connection when no channels are left after closing the final one. It defaults to infinity. - **[max_initial_idle_time](`t:ssh:max_initial_idle_time_daemon_option/0`)** - Sets a time-out on a connection that will expire if no channel is opened on the connection. The timeout is started when the authentication phase is completed. It defaults to infinity. - **[alive](`t:ssh:alive_common_option/0`)** - Sets the interval and the maximum number of alive messages that may be sent without receiving any message back. Alive messages are typically used to detect that a connection became unresponsive. The following table clarifies when a timeout is started and when it triggers: | # | Event | Timeout started | Timeout ended | |---|-------|-----------------|---------------| | 1 | TCP connected | `hello_timeout`, `negotiation_timeout` | | | 2 | First SSH message received | | `hello_timeout` | | 3 | Key Exchange finished | | | | 4 | Authenticated | `max_initial_idle_time` | `negotiation_timeout` | | 5 | Channel 1 opened | | `max_initial_idle_time` | | 6 | Channel *n* opened | | | | 7 | Channel *x_1* closed | | | | 8 | Channel *x_n* closed (all channels closed) | `idle_time` | | | 9 | Connection closed | | `idle_time` | ### Compression SSH supports compression of the data stream. Reasonable finite [max_sessions](`m:ssh#hardening_daemon_options-max_sessions`) option is highly recommended if compression is used to prevent excessive resource usage by the compression library. See [Counters and parallelism](#counters-and-parallelism). The `'zlib@openssh.com'` algorithm is recommended because it only activates after successful authentication. The `'zlib'` algorithm (deprecated in SSH and it's usage is scheduled for removal in OTP 30.0) is not recommended because it activates before authentication completes, allowing unauthenticated clients to expose potential vulnerabilities in compression libraries, and increases attack surface of compression-based side-channel and traffic-analysis attacks. In both algorithms decompression is protected by a size limit that prevents excessive memory consumption. ## Reducing Attack Surface ### Shell and Exec Services A daemon has two services for evaluating tasks on behalf of a remote client. The _exec_ server-side service takes a string provided by the client, evaluates it and returns the result. The _shell_ function enables the client to open a shell in the shell host. The options [exec](`t:ssh:exec_daemon_option/0`) and [shell](`t:ssh:shell_daemon_option/0`) are disabled per default. The same options could also install handlers for the string(s) passed from the client to the server. ### SFTP Subsystem The SFTP subsystem is not enabled by default. When enabled, SFTP provides access to the file system with the rights of the OS process running the Erlang emulator, regardless of the authenticated SSH user. See the [Terminology](terminology.md) section for details. The [subsystems](`t:ssh:subsystem_daemon_option/0`) option controls which subsystems are available. To enable SFTP: ```erlang ssh:daemon({192, 168, 1, 10}, Port, [{subsystems, [ssh_sftpd:subsystem_spec([])]} | Options]). ``` **Root directory isolation** The `root` option (see `m:ssh_sftpd`) restricts SFTP users to a specific directory tree, preventing access to files outside that directory. ```erlang ssh:daemon(Port, [ {subsystems, [ssh_sftpd:subsystem_spec([{root, "/home/sftpuser"}])]}, ... ]). ``` The `root` option is configured per daemon, not per user. All users connecting to the same daemon share the same root directory. For per-user isolation, consider running separate daemon instances on different ports or using OS-level mechanisms (PAM chroot, containers, file permissions). For high-security deployments, combine the `root` option with OS-level isolation mechanisms such as chroot jails, containers, or mandatory access control (SELinux, AppArmor). **Resource limits** When enabling the SFTP subsystem via `ssh_sftpd:subsystem_spec/1`, additional resource limits can be configured to protect against resource exhaustion attacks: `max_handles` limits the maximum number of file and directory handles that can be opened simultaneously per SFTP connection. The default is 1000. Recommended values by deployment type: - **High-security/restricted environments**: 100-256 - Minimal attack surface - Suitable for simple file transfers - May impact batch operations - **Standard production**: 500-1000 - Balances security and functionality - Supports most legitimate use cases - Recommended for general deployments - Note: The default value is 1000 - **High-throughput automation**: 1000-2000 - For backup systems, CI/CD pipelines - Parallel file operations - Monitor actual usage before increasing `max_path` limits the maximum path length accepted from SFTP clients. The default is 4096 bytes. Recommended values: - **Standard**: 4096 (default) - Accommodates most filesystem limits - Linux: typically 4096 bytes - Windows: 260 characters (legacy), 32767 (extended) - **Restricted environments**: 1024-2048 - If application uses shorter paths - Additional defense layer - Verify compatibility first `max_files` limits the number of filenames returned per READDIR request. The default is 0 (unlimited). This option prevents memory exhaustion from large directory listings. Unlike max_handles and max_path, this is primarily a performance/memory protection rather than security mitigation. Recommended values: - **Standard**: 0 (unlimited, default) - No artificial restrictions - Client handles pagination - **Large directories**: 1000-10000 - Prevents memory spikes - Improves response time - **Memory-constrained systems**: 100-1000 - Embedded systems - Resource-limited containers **Example configuration** ```erlang ssh:daemon(Port, [ {system_dir, "/etc/ssh"}, {subsystems, [ ssh_sftpd:subsystem_spec([ {root, "/sftp/chroot"}, {max_handles, 256}, {max_path, 4096}, {max_files, 1000} ]) ]}, {max_sessions, 10} ]). ``` See `m:ssh_sftpd` for complete documentation of subsystem options. ### The ID String One way to reduce the risk of intrusion is to not convey which software and which version the intruder is connected to. This limits the risk of an intruder exploiting known faults or peculiarities learned by reading the public code. Each SSH client or daemon presents themselves to each other with brand and version. This may look like ```text SSH-2.0-Erlang/4.10 ``` or ```text SSH-2.0-OpenSSH_7.6p1 Ubuntu-4ubuntu0.3 ``` This brand and version may be changed with the option [id_string](`t:ssh:id_string_common_option/0`). We start a daemon with that option: ```erlang ssh:daemon({192, 168, 1, 10}, 1234, [{id_string,"hi there"}, ... ]). ``` and the daemon will present itself as: ```text SSH-2.0-hi there ``` It is possible to replace the string with one randomly generated for each connection attempt. See the reference manual for [id_string](`t:ssh:id_string_common_option/0`). ## Cryptographic Hardening ### Algorithm Selection One of the cornerstones of security in SSH is cryptography. The development in crypto analysis is fast, and yesterday's secure algorithms are unsafe today. Therefore some algorithms are no longer enabled by default and that group grows with time. See the [SSH (App)](ssh_app.md#supported-specifications-and-standards) for a list of supported and disabled algorithms. In the User's Guide the chapter [Configuring algorithms in SSH](configure_algos.md) describes the options for enabling or disabling algorithms - [preferred_algorithms](`t:ssh:preferred_algorithms_common_option/0`) and [modify_algorithms](`t:ssh:modify_algorithms_common_option/0`). ### Post-Quantum Key Exchange The `mlkem768x25519-sha256` algorithm provides quantum-resistant key exchange using a hybrid construction that combines ML-KEM-768 (NIST FIPS 203) with X25519. When the underlying crypto library supports ML-KEM, this algorithm is negotiated by default as the highest-priority key exchange method. No configuration is needed to enable it. ### Re-Keying In the setup of the SSH connection a secret cipher key is generated by co-operation of the client and the server. Keeping this key secret is crucial for keeping the communication secret. As time passes and encrypted messages are exchanged, the probability that a listener could guess that key increases. The SSH protocol therefore has a special operation defined - _key re-negotiation_ or _re-keying_. Any side (client or server) could initiate the re-keying and the result is a new cipher key. The result is that the eavesdropper has to restart its evil and dirty craftsmanship. See the option [rekey_limit](`t:ssh:rekey_limit_common_option/0`) for a description. ## Verifying the Remote Daemon (Server) in an SSH Client Every SSH server presents a public key - the _host key_ \- to the client while keeping the corresponding private key in relatively safe privacy. The client checks that the host that presented the public key also possesses the private key of the key-pair. That check is part of the SSH protocol. But how can the client know that the host _really_ is the one that it tried to connect to and not an evil one impersonating the expected one using its own valid key-pair? There are two alternatives available with the default key handling plugin `m:ssh_file`. The alternatives are: - **Pre-store the host key** - For the default handler ssh_file, store the valid host keys in the file [`known_hosts`](`m:ssh_file#FILE-known_hosts`) and set the option [silently_accept_hosts](`m:ssh#hardening_client_options-silently_accept_hosts`) to `false`. Alternatively, write a specialized key handler using the [SSH client key API](`m:ssh_client_key_api`) that accesses the pre-shared key in some other way. - **Pre-store the "fingerprint" (checksum) of the host key** - Use [silently_accept_hosts](`m:ssh#hardening_client_options-silently_accept_hosts`) with a callback: [`accept_callback()`](`t:ssh:accept_callback/0`) or [`{HashAlgoSpec, accept_callback()}`](`t:ssh:accept_hosts/0`). ## Verifying the Remote Client in a Daemon (Server) - **Password checking** - The default password checking is with the list in the [user_passwords](`m:ssh#option-user_passwords`) option in the SSH daemon. It could be replaced with a [pwdfun](`m:ssh#option-pwdfun`) plugin. The arity four variant ([`pwdfun_4()`](`t:ssh:pwdfun_4/0`)) can also be used for introducing delays after failed password checking attempts. Here is a simple example of such a pwdfun: ```erlang fun(User, Password, _PeerAddress, State) -> case lists:member({User,Password}, my_user_pwds()) of true -> {true, undefined}; % Reset delay time false when State == undefined -> timer:sleep(1000), {false, 2000}; % Next delay is 2000 ms false when is_integer(State) -> timer:sleep(State), {false, 2*State} % Double the delay for each failure end end. ``` If a public key is used for logging in, there is normally no checking of the user name. It could be enabled by setting the option [`pk_check_user`](`m:ssh#option-pk_check_user`) to `true`. In that case the pwdfun will get the atom `pubkey` in the password argument. ## Client Connection Options A client could limit the time for the initial tcp connection establishment with the option [connect_timeout](`t:ssh:connect_timeout_client_option/0`). The time is in milliseconds, and the initial value is infinity. The negotiation (session setup time) time can be limited with the _parameter_ `NegotiationTimeout` in a call establishing an ssh session, for example `ssh:connect/3`. ## Network-Level Security ### IP Binding Restrictions `ssh:daemon/1` and `ssh:daemon/2` bind to **all network interfaces** by default. For hardened deployments, use `ssh:daemon/3` with an explicit IP address or `loopback`: ```erlang ssh:daemon({192, 168, 1, 10}, 2222, Options). % Specific interface ssh:daemon(loopback, 2222, Options). % Localhost only ``` **Note**: In the examples above, `HostAddress` (1st argument) takes precedence over a potentially provided `{ip, Address}` in `Options` (3rd argument). ## Advanced Authentication The following techniques provide enhanced authentication controls using custom callbacks. These require implementation specific to your environment. ### Public Key Validation Use a custom [`key_cb`](`t:ssh:key_cb_common_option/0`) module implementing the `m:ssh_server_key_api` behaviour. The `is_auth_key` callback can enforce client key strength requirements (e.g. reject RSA keys shorter than 2048 bits) and log key usage for auditing. Enable [`pk_check_user`](`m:ssh#option-pk_check_user`) to verify that the username is known before accepting public key authentication. ```erlang ssh:daemon({192, 168, 1, 10}, Port, [ {key_cb, {my_key_handler, []}}, {pk_check_user, true} ]). ``` ### Account Lockout Policies Use [`pwdfun`](`t:ssh:pwdfun_4/0`) with an ETS table to track failed attempts across connections and lock accounts after repeated failures. Return `disconnect` when the account is locked — this immediately terminates the connection with `SSH_DISCONNECT_NO_MORE_AUTH_METHODS_AVAILABLE`: > #### Note {: .info } > > The lockout example below is conceptual. With > [`parallel_login`](`m:ssh#hardening_daemon_options-parallel_login`) enabled, > race conditions may reduce lockout accuracy. ```erlang lockout_pwdfun(User, Password, _PeerAddr, State) -> case ets:lookup(ssh_lockouts, {locked, User}) of [_] -> disconnect; [] -> case validate_password(User, Password) of true -> ets:delete(ssh_lockouts, {attempts, User}), {true, State}; false -> N = ets:update_counter(ssh_lockouts, {attempts, User}, 1, {{attempts, User}, 0}), case N >= ?LOCKOUT_THRESHOLD of true -> ets:insert(ssh_lockouts, {{locked, User}, true}), ets:delete(ssh_lockouts, {attempts, User}); false -> ok end, {false, State} end end. ```