Writing shell scripts leaves a lot of room to make mistakes, in ways that will cause your scripts to break on certain input, or (if some input is untrusted) open up security vulnerabilities. Here are some tips on how to make your shell scripts safer.

Don't

The simplest step is to avoid using shell at all. Many higher-level languages are both easier to write the code in in the first place, and avoid some of the issues that shell has. For example, Python will automatically error out if you try to read from an uninitialized variable (though not if you try to write to one), or if some function call you make produces an error.

One of shell's chief advantages is that it's easy to call out to the huge variety of command-line utilities available. Much of that functionality will be available through libraries in Python or other languages. For the handful of things that aren't, you can still call external programs. In Python, the subprocess module is very useful for this. You should try to avoid passing shell=True to subprocess (or using os.system or similar functions at all), since that will run a shell, exposing you to many of the same issues as plain shell has. It also has two big advantages over shell — it's a lot easier to avoid word-splitting or similar issues, and since calls to subprocess will tend to be relatively uncommon, it's easy to scrutinize them especially hard. When using subprocess or similar tools, you should still be aware of the suggestions in "Passing filenames or other positional arguments to commands" below.

Shell settings

POSIX sh and especially bash have a number of settings that can help write safe shell scripts.

I recommend the following in bash scripts:

set -euf -o pipefail

In dash, set -o doesn't exist, so use only set -euf.

What do those do?

set -e

If a command fails, set -e will make the whole script exit, instead of just resuming on the next line. If you have commands that can fail without it being an issue, you can append || true or || : to suppress this behavior — for example set -e followed by false || : will not cause your script to terminate.

set -u

Treat unset variables as an error, and immediately exit.

set -f

Disable filename expansion (globbing) upon seeing *, ?, etc..

If your script depends on globbing, you obviously shouldn't set this. Instead, you may find shopt -s failglob useful, which causes globs that don't get expanded to cause errors, rather than getting passed to the command with the * intact.

set -o pipefail

set -o pipefail causes a pipeline (for example, curl -s https://sipb.mit.edu/ | grep foo) to produce a failure return code if any command errors. Normally, pipelines only return a failure if the last command errors. In combination with set -e, this will make your script exit if any command in a pipeline errors.

Quote liberally

Whenever you pass a variable to a command, you should probably quote it. Otherwise, the shell will perform word-splitting and globbing, which is likely not what you want.

For example, consider the following:

alex@kronborg tmp [15:23] $ dir="foo bar"
alex@kronborg tmp [15:23] $ ls $dir
ls: cannot access foo: No such file or directory
ls: cannot access bar: No such file or directory
alex@kronborg tmp [15:23] $ cd "$dir"
alex@kronborg foo bar [15:25] $ file=*.txt
alex@kronborg foo bar [15:26] $ echo $file
bar.txt foo.txt
alex@kronborg foo bar [15:26] $ echo "$file"
*.txt

Depending on what you are doing in your script, it is likely that the word-splitting and globbing shown above are not what you expected to have happen. By using "$foo" to access the contents of the foo variable instead of just $foo, this problem does not arise.

When writing a wrapper script, you may wish pass along all the arguments your script received. Do that with:

wrapped-command "$@"

See "Special Parameters" in the bash manual for details on the distinction between $*, $@, and "$@" — the first and second are rarely what you want in a safe shell script.

Passing filenames or other positional arguments to commands

If you get filenames from the user or from shell globbing, or any other kind of positional arguments, you should be aware that those could start with a "-". Even if you quote correctly, this may still act differently from what you intended. For example, consider a script that allows somebody to run commands as nobody (exposed over remctl, perhaps), consisting of just sudo -u nobody "$@". The quoting is fine, but if a user passes -u root reboot, sudo will catch the second -u and run it as root.

Fixing this depends on what command you're running.

For many commands, however, -- is accepted to indicate that any options are done, and future arguments should be parsed as positional parameters — even if they look like options. In the sudo example above, sudo -u nobody -- "$@" would avoid this attack (though obviously specifying in the sudo configuration that commands can only be run as nobody is also a good idea).

Another approach is to prefix each filename with ./, if the filenames are expected to be in the current directory.

Temporary files

TODO: mumble mktemp?

Use ShellCheck to check for bugs

The ShellCheck linter automatically catches a number of the above mistakes and more. Run it regularly, ideally with integration into your editor and your test suite, and address all of its diagnostics. Even warnings that might sound unimportant could be obscuring important bugs.

Other resources

Google has a Shell Style Guide. As the name suggests, it primarily focuses on good style, but some items are safety/security-relevant.

Conclusion

When possible, instead of writing a "safe" shell script, use a higher-level language like Python. If you can't do that, the shell has several options that you can enable that will reduce your chances of having bugs, and you should be sure to quote liberally.