Credit: alegrey91 for the logo! Thanks!
Pakkero is a binary packer written in Go made for fun and educational purpose.
Its main goal is to take in input a program file (elf binary, script, even appimage) and compress it, protect it from tampering and intrusion.
It is not recommended for very small files as the launcher itself can vary from ~700kb to ~1.7mb depending on compression. On files above 2.6mb there is gain, else the resulting binary is larger than the original:
base-bin 1.2M -> 1.6M
smaller-bin 2.4M -> 2.3M
small-bin 3.7M -> 3.0M
medium-bin 25M -> 16M
big-bin 148M -> 88M
With compression disabled, all resulting file size are ~1mb higher, making it suitable for 5+mb files.
Tested with a 24mb binary file (I didn't have a big project handy so I just concatenated a bunch of programs from /usr/bin/ to make one big elf) became:
- 12mb using
upx -9 - 13mb using
pakkero -c - 14mb using
pakkerowithout compression
If you have a Go environment ready to go, it's as easy as:
go get github.com/89luca89/pakkeroOnce you retrieved you are ready to build:
cd $GOPATH/src/github.com/89luca89/pakkero; makeor to test
cd $GOPATH/src/github.com/89luca89/pakkero; make testThe binary file will be in $GOPATH/src/github.com/89luca89/pakkero/dist
The following are hard dependencies:
- go -> to build the launcher
- ls
- sed
- strip -> to strip the launcher
The following are weak dependencies
- upx -> needed for launcher compression (optional)
GO 1.13+ needed
Dependencies are checked at runtime and an error message will specify what is missing
This is a for-fun and educational project, complete protection for a binary is impossible, in a way or another there is always someone that will reverse it, even if only based on 0 an 1, so this is more about exploring some arguments that to create an anti-reverse launcher.
Pakkero is divided in two main pieces, the packer part (Pakkero itself) and the launcher part.
Pakkero can be launched like:
pakkero --file ./target-file -o ./output-file -register-dep dependency-file -c
Typing pakker -h the following output will be shown:
Usage: pakkero -file /path/to/file -offset OFFSET (-o /path/to/output) (-c) (-register-dep /path/to/file)
-file <file> Target file to Pack
-o <file> place the output into <file> (default is <inputfile>.enc), optional
-c compress the output to occupy less space (uses UPX), optional
-offset Offset where to start the payload (Number of Bytes)
-enable-stdout Whether to wait and handle the process stdout/sterr or not (false by default, optional)
-register-dep /path/to/dependency to analyze and use as fingerprint (absolutea, optional)
-v Check pakkero versionBelow there is a full explanation of provided arguments:
- file: The file we want to pack
- o: (optional) The file output that we will create
- c: (optional) If specified, UPX will be used to further compress the Launcher
- offset: (optional) The number of bytes from where to start the payload (increases if not using compression)
- enable-stdout (optional) whether to enable or not the handling of the stdout/err of the payload disabled by default, less secure
- regiser-dep (optional) Path to a file that can be used to register the fingerprint of a dependency to ensure that the Launcher runs only if a file with similar fingerprint is present
- v: Print version
The main intent is to not alter the payload in any way, this can be very important for types of binary that rely on specific order of instructions or relatively fragile timings.
The target of pakkero is not to touch the "payload". Other packers like UPX works by compressing the sections stored within the Section Table of the executable file, relocating the sections and renaming them. It then alters the entry point where the binary will run. While this is really what defines storically a packer, this in some way or another "touches" the payload so can make it unusable (when it works on strict timing, precise elf sections tricks and so on)
To build the project, you can simply use the Makefile;
-
makewill compile -
make testwill compile and run a run with a simple binary (echo)
Why not using simply go build?
Go build works fine, but will skip a fundamental step in the building process, the injection of the launcher stub inside Pakkero source
This way the Pakkero binary has inside the source of the Launcher to be used for each packaging.
Build Pakkero image:
sudo docker build . -t pakkeroRun containerized Pakkero:
sudo docker run -it -v <target_dir>:/ext pakkero --file /ext/<target_elf> -o /ext/<target_elf>.packedBuild Pakkero image:
sudo podman build . -t pakkeroRun containerized Pakkero:
sudo podman run -it -v <target_dir>:/ext pakkero --file /ext/<target_elf> -o /ext/<target_elf>.packedFor this purpose the payload is simply compressed using zlib then encrypted using AES256-GCM
During encryption, some basic operations are also performed on the payload:
- putting garbage random values before and after the payload to mask it
- reverse it and change each byte endianess
Encryption password is the hash SHA512 of the compiled launcher itself together with the garbage values added to fill the file till the offset, thus providing some integrity protection and anti-tampering.
The offset will decide where in the output file the payload starts.
Put simply, after the launcher is compiled (more on the launcher later), the payload is attached to it. The offset ensures that the payload can be put anywhere after it. All the space after the launcher until the payload is filled with random garbage.
Being part of the password itself, greater offset will make stronger the encryption, but enlarge the final output file.
Optimal value are at least 800000 when compression is enabled and 1900000 when disabled. *If not specified a random one will be chosen upon creation.
The final thing the packer does is compiling the launcher. To protect some of the fundamental part of it (namely where the offset starts) the launcher is obfuscated and heavily stripped down.
The technique utilized for obfuscating the function and variables name is based on typo-squatting:
This is done in a pretty naive way, simply put, in the launcher each function/variable which name has to be obfuscated, needs to start with the suffix ob, it will be then put into a secret map, and each occurrence will be replaced in the file with a random string of length 128, composed only of runes that have similar shape, namely:
mixedRunes := []rune("0OÓÕÔÒÖŌŎŐƠΘΟ")For pure strings in the launcher, they are detected using regular expressions, finding all the words that are comprised between the three type of ticks supported in go
`
'
"
All of the strings found this way, are then replaced with a function that performs a simple operation of reconstruction of the original string:
func ÓΘŌOÒŐÒŌÓÒOŎΘOΟ0ŐÒÖŎÕΟΘÕÓÕÓŎÓŌÕ0ŎŌΘÕŎÕ() string {
ŌÒ0ŎŎŐÓÖÖΘO0ŌŌŌÒŌŌÒƠÔÖΘŐÖΟŎƠƠ00Õ0ÖÕ0ÖŐŐÓΟŌΟ := []string{"Ò0ƠŐÖŐŎΘÖƠÔÖÕÓΘÕÕŌŎŐƠÔΘΘƠ", "ÔÒÕΟƠŐÒŌOƠÖ", "ƠΘŐÒƠΘŌ00ΘΘΟÔŎŎΘŐƠŐΘŎΟÕÖÕÖΟÖΘÒÖ"}
var ΟŐÔÒÖÔΘÕÔŎÒÓÖÖÒΘ0ÖÔΟÖŎ0ÔOÓÖƠŌÔÓŌŌ []byte
for _, ŌΘƠÔÕÔΘÔOÔOŐΟΘŌΘƠÔÕÔΘÔOÔOŐΟΘ := range ŌÒ0ŎŎŐÓÖÖΘO0ŌŌŌÒŌÕÒÔÕŌŎ00ÔÔÒOƠÓÕÔÒ0ΘƠΘŐ0OOŎÓÒŐ0Õ0ÓOÕÓŐƠŌŎÕÖ0ÕÖÔŌΟΟŌÒÖÒÖOΟOÒ0ÖŐŐÓΟŌΟ {
ΟŐÔÒÖÔΘÕÔŎÒÓÖÖÒΘ0ÖÔΟÖŎ0ÔOÓÖƠŌÔÓŌŌ = append(ΟŐÔÒÖÔΘÕÔŎÒÓÖÖÒΘ0ÖÔΟÖŎ0ÔOÓÖƠŌÔÓŌŌ, byte(len([]rune(ŌΘƠÔÕÔΘÔOÔOŐΟΘŌΘƠÔÕÔΘÔOÔOŐΟΘ))))
}
return string(ΟŐÔÒÖÔΘÕÔŎÒÓÖÖÒΘ0ÖÔΟÖŎ0ÔOÓÖƠŌÔÓŌŌ)
}A slice of string is generated, with each element has a lenght derived from the byte value of the original char of the original string. This way each byte of the original string is computed and calculated as the lenght of the correspondent string, casted to rune slice.
The launcher is compiled then using:
flags = []string{"build", "-a",
"-trimpath",
"-gcflags",
"-N -l -nolocalimports",
"-ldflags",
"-s -w -extldflags -static",
}
exec.Command("go", flags...)File is the stripped, using strip with the flags:
-sxX
--remove-section=.bss
--remove-section=.comment
--remove-section=.eh_frame
--remove-section=.eh_frame_hdr
--remove-section=.fini
--remove-section=.fini_array
--remove-section=.gnu.build.attributes
--remove-section=.gnu.hash
--remove-section=.gnu.version
--remove-section=.gosymtab
--remove-section=.got
--remove-section=.note.ABI-tag
--remove-section=.note.gnu.build-id
--remove-section=.note.go.buildid
--remove-section=.shstrtab
--remove-section=.typelinkAdditionally, if using UPX, their headers are removed and replaced with randomness, to ensure simple things like upx -d will not work.
Additionally a series of extra words are removed from the binary and replaced with random bytes, to make it harder to do static analysis:
.gopclntab
.go.buildinfo
.noptrdata
.noptrbss
.data
.rodata
.text
.itablink
.shstrtab
.data
.dynamic
.dynstr
.dynsym
.gnu.version_r
.gopclntab
.got.plt
.init_array
.interp
.itablink
.rela.dyn
.rela.plt
.tbss
.plt
.init
name runtime command cmd
ptr process unicode main
path get reflect context
debug fmt sync sort
size heap fatal call
fixed slice bit file
read write buffer encrypt
decrypt hash state
external internal float
env trace pid
Output of readelf to see the effect:
Using binwalk to analyze the file entropy can give some hint on how the process works:
This is the entropy of the binary we want to package (for this example /usr/bin/bash):
This is the entropy of a packaged binary without compression
This is the entropy of a packaged binary with compression
In both cases (but mainly the first) it is possible to see when the launcher stops and the garbage before the payload starts. This is really not a problem, because the offset of garbage is both pre-poned and post-poned to the payload, and the "secret number" of when it starts is kept inside the launcher and computed at runtime.
This is obviously vulnerable, reversing the binary will reveal the secret, all the launcher part is dedicated to the implementation of a series of measures to block dynamic analysis and try to force static analysis.
The launcher is the second part of the project, it allows to decompress, decrypt and launch the payload without touching storage, but using a file descriptor in RAM.
This is a well known technique as it is possible to read:
and in many other places in C programming literature.
Put briefly, use syscall to create a memory file descriptor (syscall 319 for amd64),
write the plaintext payload here, and execute.
The fd will be automatically removed after the execution without
leaving trace on the storage.
This approach is vulnerable to
-
"memory dump attack", for example pausing the VM during execution and manually search the ram for all file descriptors until you find the right one
-
"proc dump attack", in linux all the file descriptors are in
/procso dumping to another disk the complete folder will in a way or another dump the decrypted payload (if done before the execution finishes), this is even more accentuated if stdout management is enabled -
dynamic analysis, during execution "pausing" the process and spot the right fd
-
reversing the binary to find the "secret" (which in our case is the offset) and from there, reverse the encryption process and reconstruct the plaintext
For point 1, it is possible to insert hypervisor detection, sandbox detection etc... it is in my TODO list, but I would leave it optional, in case you genuinely want to run the binary in VMs or Dockers.
Point 3 (and by consequence point 4) can be made harder by blocking dynamic analysis detecting debuggers, tracers and so on...
Forcing static analysis of the decompiled code is already a big step forward in protecting the binary execution.
Implemented here are a series of anti-debug techniques that are quite common in C/C++, from the double-ptrace method to the ppid analysis and breakpoints interception.
First line of protection is breakpoints interception, on linux, a breakpoints is equivalent to signal SIGILL and SIGTRAP so:
/*
Breakpoint on linux are 0xCC and will be interpreted as a
SIGTRAP, we will intercept them.
*/
func obSigTrap(obInput chan obOS.Signal) {
obMySignal := <-obInput
switch obMySignal {
case obSyscall.SIGILL:
obExit()
case obSyscall.SIGTRAP:
obExit()
default:
return
}
}This is pretty basic, so we go ahead and try to block ptrace:
// attach to PTRACE, register if successful
// attach A G A I N , register if unsuccessful
// this protects against custom ptrace (always returning 0)
// against NOP attacks and LD_PRELOAD attacks
func obPtraceDetect() {
var obOffset = 0
obProc, _ := obOS.FindProcess(obOS.Getppid())
obErr := obSyscall.PtraceAttach(obProc.Pid)
if obErr == nil {
obOffset = 5
}
obErr = obSyscall.PtraceAttach(obProc.Pid)
if obErr != nil {
obOffset *= 3
}
if obOffset != (3 * 5) {
obProc.Signal(obSyscall.SIGCONT)
println(1)
return
}
obErr = obSyscall.PtraceDetach(obProc.Pid)
if obErr != nil {
obProc.Signal(obSyscall.SIGCONT)
println(0)
return
}
obProc.Signal(obSyscall.SIGCONT)
println(0)
}Double ptraceme ensures that tampering the ptrace loading with a fake ptrace lib that always returns 0, would result in a failure.
CMD Line detection, would check for common processes for debugging, this is pretty naive check:
/*
Check the process cmdline to spot if a debugger is inline
*/
func obParentCmdLineDetect() {
obPidParent := obOS.Getppid()
obNameFile := "/proc/" + obStrconv.FormatInt(int64(obPidParent), 10) +
"/cmdline"
obStatParent, _ := obUtilio.ReadFile(obNameFile)
if obStrings.Contains(string(obStatParent), "gdb") ||
obStrings.Contains(string(obStatParent), "dlv") ||
obStrings.Contains(string(obStatParent), "edb") ||
obStrings.Contains(string(obStatParent), "frida") ||
obStrings.Contains(string(obStatParent), "ghidra") ||
obStrings.Contains(string(obStatParent), "godebug") ||
obStrings.Contains(string(obStatParent), "ida") ||
obStrings.Contains(string(obStatParent), "lldb") ||
obStrings.Contains(string(obStatParent), "ltrace") ||
obStrings.Contains(string(obStatParent), "strace") ||
obStrings.Contains(string(obStatParent), "valgrind") {
obExit()
}
}/*
Check the process cmdline to spot if a debugger is the PPID of our process
*/
func obParentDetect() {
obPidParent := obOS.Getppid()
obNameFile := "/proc/" + obStrconv.FormatInt(int64(obPidParent), 10) +
"/stat"
obStatParent, _ := obUtilio.ReadFile(obNameFile)
if obStrings.Contains(string(obStatParent), "gdb") ||
obStrings.Contains(string(obStatParent), "dlv") ||
obStrings.Contains(string(obStatParent), "edb") ||
obStrings.Contains(string(obStatParent), "frida") ||
obStrings.Contains(string(obStatParent), "ghidra") ||
obStrings.Contains(string(obStatParent), "godebug") ||
obStrings.Contains(string(obStatParent), "ida") ||
obStrings.Contains(string(obStatParent), "lldb") ||
obStrings.Contains(string(obStatParent), "ltrace") ||
obStrings.Contains(string(obStatParent), "strace") ||
obStrings.Contains(string(obStatParent), "valgrind") {
obExit()
}
}this goes in conjunction with the TracePid check to see if a parent is tracing us:
/*
Check the process status to spot if a debugger is active using the TracePid key
*/
func obParentTracerDetect() {
obPidParent := obOS.Getppid()
obNameFile := "/proc/" + obStrconv.FormatInt(int64(obPidParent), 10) +
"/status"
obStatParent, _ := obUtilio.ReadFile(obNameFile)
obStatLines := obStrings.Split(string(obStatParent), "\n")
for _, obValue := range obStatLines {
if obStrings.Contains(obValue, "TracerPid") {
obSplitArray := obStrings.Split(obValue, ":")
obSplitValue := obStrings.Replace(obSplitArray[1], "\t", "", -1)
if obSplitValue != "0" {
obExit()
}
}
}
}and verification that the process cmdline corresponds to "argv[0]" (for example, launching strace mybin arg1 would result in a cmdline of strace and argv[0] of mybin
/*
Check the process cmdline to spot if a debugger is launcher
"_" and Args[0] should match otherwise
*/
func obEnvArgsDetect() {
obLines, _ := obOS.LookupEnv("_")
if obLines != obOS.Args[0] {
obExit()
}
}and check if there is none of the known debuggers inline with the command
/*
Check the process cmdline to spot if a debugger is inline
"_" should not contain the name of any debugger
*/
func obEnvParentDetect() {
obLines, _ := obOS.LookupEnv("_")
if obStrings.Contains(obLines, "gdb") ||
obStrings.Contains(obLines, "dlv") ||
obStrings.Contains(obLines, "edb") ||
obStrings.Contains(obLines, "frida") ||
obStrings.Contains(obLines, "ghidra") ||
obStrings.Contains(obLines, "godebug") ||
obStrings.Contains(obLines, "ida") ||
obStrings.Contains(obLines, "lldb") ||
obStrings.Contains(obLines, "ltrace") ||
obStrings.Contains(obLines, "strace") ||
obStrings.Contains(obLines, "valgrind") {
obExit()
}
}Also a pretty common check to do if we are in a debugger is to see if ENV has the variables LINES and COLUMNS:
/*
Check the process cmdline to spot if a debugger is active
most debuggers (like GDB) will set LINE,COLUMNS or LD_PRELOAD
to function, we try to spot this
*/
func obEnvDetect() {
_, obLines := obOS.LookupEnv("LINES")
_, obColumns := obOS.LookupEnv("COLUMNS")
_, obLineLdPreload := obOS.LookupEnv("LD_PRELOAD")
if obLines || obColumns || obLineLdPreload {
obExit()
}
}This can give false positives (example the embedded terminal in VSCode or VIM), but worth checking for a normal environment.
Also banally checking for LD_PRELOAD env variable present:
/*
Check the process is launcher with a LD_PRELOAD set.
This can be an injection attack (like on frida) to try and circumvent
various restrictions (like ptrace checks)
*/
func obLdPreloadDetect() {
obKey := obStrconv.FormatInt(obTime.Now().UnixNano(), 10)
obValue := obStrconv.FormatInt(obTime.Now().UnixNano(), 10)
err := obOS.Setenv(obKey, obValue)
if err != nil {
obExit()
}
obLineLdPreload, _ := obOS.LookupEnv(obKey)
if obLineLdPreload == obValue {
err := obOS.Unsetenv(obKey)
if err != nil {
obExit()
}
} else {
obExit()
}
}to make it more resilient to "false environment" attacks, we also try and set a random key-value in the environment, and check if it works, to ensure we do not have a "fake" environment (always empty for example).
This type of checks are pretty basic and easy to port from C to Go.
A couple of checks I would like to port are for example the heap relocation check, as explained in this repo: debugmenot/test_nearheap.c at master · kirschju/debugmenot · GitHub
GDB relocates the heap to the end of the bss section
This type of check is not easily done in Go because go does not support pointer arithmetic, CGO should be the way, but would make it dynamically linked for the C part (or twice the size if statically linked)
To add to this in many points of the source code it is possible to see that a comment is made:
//OB_CHECK
This is because during the obfuscation phase, every time we encounter this string, we will inject a random permutation of ALL of the anti-debug tricks that will be executed in parallel using go routines.
/*
GenerateRandomAntiDebug will Insert random order of anti-debug check
together with inline compilation to induce big number
of instructions in random order
*/
func GenerateRandomAntiDebug(input string) string {
lines := strings.Split(input, "\n")
randomChecks := []string{
`obDependencyCheck()`,
`obEnvArgsDetect()`,
`obParentTracerDetect()`,
`obParentCmdLineDetect()`,
`obEnvDetect()`,
`obEnvParentDetect() `,
`obLdPreloadDetect()`,
`obParentDetect()`,
}
// find OB_CHECK and put the checks there.
for i, v := range lines {
if strings.Contains(v, "// OB_CHECK") {
threadString := ""
checkString := ""
// randomize order of check to replace
for j, v := range ShuffleSlice(randomChecks) {
threadString = threadString + "go " + v + ";"
checkString += v
if j != (len(randomChecks) - 1) {
checkString += `||`
}
}
// add action in case of failed check
lines[i] = threadString
}
}
// back to single string
return strings.Join(lines, "\n")
}The generated source will be filled with this always-changing code, that will make difficult to use NOP attacks, manual jump using breakpoints and make confusion in the graph view of disassemblers like Cutter, IDA Pro or Ghidra.
So the main, for example, becomes something like:
func main() {
// Prepare to intercept SIGTRAP
ÖÖΘÓOŐƠOΟŌΟŐÔÒÕÓÒÒÒOŎŌO0ÔŌÓÖOŐŎΟŌŌŌÔ0OOŌŐΘŎΘÔOOOÓΘŐΘ00ƠŎÖÖ0ÔÕÕŎŎŎÒÖO00ŎÔΘÓOÖ0ΘÔOΟÒŐŐƠÒΘÒ0ŎŌŌÖΟÓ0ƠŎŌŌΟŎΘŐÖÒŐΘÔÕOŌÕŎÒŎÔŌÓƠÓŌŐΟΟÒÒƠ := make(chan ÒÕ0ΟÓŎ0ÖÒŌƠÒΘ0ÓΘΘƠΟÕƠΘÕO0OŐOŎOƠŎƠÖOOÒÖÒO0ÕÕÓÖΘΘÓÒÕŌŐΘŐƠÖOÕΘÔÕÖƠƠŌŐƠ0ÖÖƠŌΟ0ΘƠÕÒÔƠÒOÒΘÔΟÒ0OÕOŌÕΘÒÒŐÓΟÕÖÕŎΟƠŎOŌÔΟŎŐÒŎŎOƠOO0Ō0ŌŐÓÖŎÖ.Signal, 1)
ÔƠÕΟÒΘΘÖŎ0ÓŎÒŎOÕΟÒ0ŐΘƠÓOƠOÔÒÕŐ0OÔÒÓΟOÓŐŐŎΟÒŐÒ0ŌΟO0ΟOÔOΘOƠƠÒ0ŐÓΘÖΘΘÕÖΟÔÖÖΘÔŎÕ00ŐÖÔŎŎŌ0ŐƠOÖÖΟΘÕÓŌÖÔÖÕÒƠƠÓÓ0ÒÖƠŌÔÔŐƠŌOŐÔÒ0ŐÕÖOŐŌŎÔÔ.Notify(ÖÖΘÓOŐƠOΟŌΟŐÔÒÕÓÒÒÒOŎŌO0ÔŌÓÖOŐŎΟŌŌŌÔ0OOŌŐΘŎΘÔOOOÓΘŐΘ00ƠŎÖÖ0ÔÕÕŎŎŎÒÖO00ŎÔΘÓOÖ0ΘÔOΟÒŐŐƠÒΘÒ0ŎŌŌÖΟÓ0ƠŎŌŌΟŎΘŐÖÒŐΘÔÕOŌÕŎÒŎÔŌÓƠÓŌŐΟΟÒÒƠ, O0OÓΟ0ŎŐOƠOŎƠ0ÓƠΘÒƠÕƠÖÓÕŎƠŎOΟΘƠŌƠÖÔÓÔÕŐŎŌŐŐƠΘO0Ô0Ò0OÕƠ0ŐÓÖÔ0ŎΘÕÔÔŎƠÓÔOÕO0ÓΘÒOŎŎÔ0ÓƠ0ÖŎŎΘO0ƠŐÖŐŌOOOΟOÒΘΟΟƠΘΟ0ÖÔŎƠÓŌÕÕOΘÕŎŐŌΘOŐΘÕŌ.SIGTRAP, O0OÓΟ0ŎŐOƠOŎƠ0ÓƠΘÒƠÕƠÖÓÕŎƠŎOΟΘƠŌƠÖÔÓÔÕŐŎŌŐŐƠΘO0Ô0Ò0OÕƠ0ŐÓÖÔ0ŎΘÕÔÔŎƠÓÔOÕO0ÓΘÒOŎŎÔ0ÓƠ0ÖŎŎΘO0ƠŐÖŐŌOOOΟOÒΘΟΟƠΘΟ0ÖÔŎƠÓŌÕÕOΘÕŎŐŌΘOŐΘÕŌ.SIGILL)
go ÔΘƠƠƠŎÔŐŌÖOŐÒÓO0ÖΘΘ0ΘŐŎΘ0ÖŌOÒŐÕ0ÒŐ0ƠÖÕŐÓÕÕŌƠƠÖÒÔŎÔÔÖÕŐÖΟÒÖŐÓÕÔÓΘΟÕƠOŐ0ŌÓÖÔOƠÓ0ŎÓ0ŎOÒΟÓOΘÔΟÒÔƠΘŐÒΘΘÓƠΟƠÓÒÓƠ0Ŏ0ÓÒŌOŌÕÕΟŌOΘ0ŎƠŎ0ÕƠƠ(ÖÖΘÓOŐƠOΟŌΟŐÔÒÕÓÒÒÒOŎŌO0ÔŌÓÖOŐŎΟŌŌŌÔ0OOŌŐΘŎΘÔOOOÓΘŐΘ00ƠŎÖÖ0ÔÕÕŎŎŎÒÖO00ŎÔΘÓOÖ0ΘÔOΟÒŐŐƠÒΘÒ0ŎŌŌÖΟÓ0ƠŎŌŌΟŎΘŐÖÒŐΘÔÕOŌÕŎÒŎÔŌÓƠÓŌŐΟΟÒÒƠ)
// ÔOÒŎ00ƠÒŎΟOÓΟÖΟOÖΘÓƠƠƠŎÖŌŐŐÒÓÔOŎÒÔΘŌΟÓŐÒΟƠOÕÖÕƠƠÕŎÔÔÒΟOÖO0ŐÔÓÓƠÒ0ŎΘOÒŎOŐÔŌƠÓÒÕÕΟO0ÓƠÒŎŌÖOŌÖOÕÒƠŐÓΟŐÔÔŎŐÓÔÓOƠΘOOŎ0ƠΘÓΘÕÔÕƠÓΘÒÔŐΟÒ()
go ÔŐŌŌÒÓ0OΟO0ÖΟŌ0ƠÒ0ÒŌÔOΟΟÖŎÕ0ΘÓΘOOΟΟΘ0Ơ0ΘŐÕ0ΟÔŐ0ŐŌÕ00ŌOΟΟÒÕÒÖŌÓOOÒÔΘÒΘÒÔŌŎΟÒŐŌÕÒÕŎÖÒÖÕΘ00OÖÓŌÓÖÖÓÔƠΘŌ0ÓÕ0ŌÖΘOÖ0ÓŐƠ0ΟÖ00ŌŐŐΘΘÕÒƠÕΟ()
go ÖƠŎŎΘΟÓÓÒÖÒ0ÓΘŎÖΘŌ0ŎOÕÔÓƠÔÔΘŌÔΘ0OÕOΟOÕƠŌÒŎ0O0ŎΘΟÒÕΟΟÓ0ŎΟÒΘÕΟÖŎÕÔOÕOΘΘŐÖŌŐΟÔΟÖΟ0OÔÓ0ΟŐOŌÕŌOŌOÕÒÒΘΟŎƠΟÒŌ0ƠΟÔOÒÓÒΟŐÕÒŌƠÔÖ0ŐÖÖÖÓΟƠŌÕ()
go ƠΘÔŐŎŌÒŌÖŌΟŎOΟŐŎŎΘOƠÔ0ŐƠΘÒΟƠŐŌ0ŌΘΟΟÒŎÖÓ0ÒΟŐ0ƠΘŌÒÔÔŌƠŌÖOƠÓÔŌOÓ0OŐÕÒ0ŌÖÓŎÕŎÕÓÓ0ÒΘÖΟΘŎÕÓƠŐΟÕΟŌÔŎŌÓOÒÓ0Ŏ0ΟÕÔÒŌƠΘŌΟŎŐΘƠÔÓŐΘŎÔŌŌO0ƠÒÖÖ()
go ΟƠŐΟÔOƠÔΟΟÓΘÖOŐΟÕƠ0ΘŐÓÖÔŎÒ0OÔŐÓΟƠΘΘΘÔOÖŌÕŎÒŎΟÔÕÒŎÓ0OÖƠ0ÖÓOÒŐÓŐÔΘÓŌƠÕŎÕÒΟÖŌÒŎΟ0ÒÔÔÕ0ÕÓÒÓÒÔ0ŎO0Θ0ΘŐÖΟŎÖŌOΘÒΘ0ÕŐÖÓΟOÖÒÒÔÕOƠÒŎΘO0ÔÓÔ()
go ÒŎ0ÖƠÔŌÔÖΟΘƠΟÒƠƠÖ0ŐÓÓΟOÓŎÖΟÔŌOOΟÔÒÓƠÓΟŎŎ0ÕƠŌÕÒÒÖ0OŌÕŐ0ƠÖΟƠÕÔÒÖŎÖÔÕÔŐOƠŌƠÒΘŌŎÖOŎΘÓÔÔΟŐŌΟŐƠÕΟÖŐ0ÔΘƠŌƠƠÒOOŎÔ0ŐΟÕÕŌÕÔÕÒŐÔŐÔÖÓƠÖΟOŌŎƠ()
go ÔŌÕŌŐ0ŐÓÖΘŐÖΟŎΟƠŎOÖÖÕÒΘƠŐΟÓΟΘOÒƠΟ0ÒÒÖÖŌŎŌΘƠƠOΟÒƠΟΘOÓÕŌÕÒÕO0ŎOÔŐOÖŎÓÔOO0OÕŌOΘ0ÒƠƠÔŌÔÔÔŐΟ0ΘΘOŌOÒÕÒÕÕOÔÕŐŌΟÒŌÔ00ÖƠÔΟ0ΟÔŐΟƠÖÕΘ0ÖΟŌÖÖ()
go ÒŎÒ0ΘŐŐÔƠΘÒŐŎΘÓÕÕÔŎÖ0ŌÖÖΟÖÓÔÓÒÖÖŌŐΘŌ0ÕÖŎÓΟŌŌÒÖ0ΘÖOΘŎŎÖΘΟÓÒƠΟŐOƠOΘΘÓ0ÕΘŐƠÔÖΟΟ0OΘÖΘŎΟŐÕΘŐΟÓÓƠƠŐÒΟÖΘOŌƠΟΘŎ0ÒΟƠƠÔŎ00ŌÕÓÒ0ŌŎÒOƠŎÖŐÕÓƠ()
go ŎÓÕƠŌOŌΟÖÓ0ÒÓΘΟ0ƠŎÖΟŌŎÕÒÕÓŎÖÕƠÒŎŐƠÒƠ0ÒƠÖÕŎÒŐÓÔÓÒ0ΟÕŌÒŐOŐŌÖOÕƠƠÔŎÖ0ÒÓÒÕOÖÕƠÔÕOÔÖΟΘÕÔ0OÕƠÓOÓOÔÖÒ0ΟÖOÖÔΘŌOŌŎΘΘΘ0ŐŌOÔÒΟƠ0ŐÓÕŐÔŐΟÒŐÒΟ()
ΟƠŐΟÔOƠÔΟΟÓΘÖOŐΟÕƠ0ΘŐÓÖÔŎÒ0OÔŐÓΟƠΘΘΘÔOÖŌÕŎÒŎΟÔÕÒŎÓ0OÖƠ0ÖÓOÒŐÓŐÔΘÓŌƠÕŎÕÒΟÖŌÒŎΟ0ÒÔÔÕ0ÕÓÒÓÒÔ0ŎO0Θ0ΘŐÖΟŎÖŌOΘÒΘ0ÕŐÖÓΟOÖÒÒÔÕOƠÒŎΘO0ÔÓÔ()
go ÔŐŌŌÒÓ0OΟO0ÖΟŌ0ƠÒ0ÒŌÔOΟΟÖŎÕ0ΘÓΘOOΟΟΘ0Ơ0ΘŐÕ0ΟÔŐ0ŐŌÕ00ŌOΟΟÒÕÒÖŌÓOOÒÔΘÒΘÒÔŌŎΟÒŐŌÕÒÕŎÖÒÖÕΘ00OÖÓŌÓÖÖÓÔƠΘŌ0ÓÕ0ŌÖΘOÖ0ÓŐƠ0ΟÖ00ŌŐŐΘΘÕÒƠÕΟ()
go ƠΘÔŐŎŌÒŌÖŌΟŎOΟŐŎŎΘOƠÔ0ŐƠΘÒΟƠŐŌ0ŌΘΟΟÒŎÖÓ0ÒΟŐ0ƠΘŌÒÔÔŌƠŌÖOƠÓÔŌOÓ0OŐÕÒ0ŌÖÓŎÕŎÕÓÓ0ÒΘÖΟΘŎÕÓƠŐΟÕΟŌÔŎŌÓOÒÓ0Ŏ0ΟÕÔÒŌƠΘŌΟŎŐΘƠÔÓŐΘŎÔŌŌO0ƠÒÖÖ()
go ÖƠŎŎΘΟÓÓÒÖÒ0ÓΘŎÖΘŌ0ŎOÕÔÓƠÔÔΘŌÔΘ0OÕOΟOÕƠŌÒŎ0O0ŎΘΟÒÕΟΟÓ0ŎΟÒΘÕΟÖŎÕÔOÕOΘΘŐÖŌŐΟÔΟÖΟ0OÔÓ0ΟŐOŌÕŌOŌOÕÒÒΘΟŎƠΟÒŌ0ƠΟÔOÒÓÒΟŐÕÒŌƠÔÖ0ŐÖÖÖÓΟƠŌÕ()
go ÒŎÒ0ΘŐŐÔƠΘÒŐŎΘÓÕÕÔŎÖ0ŌÖÖΟÖÓÔÓÒÖÖŌŐΘŌ0ÕÖŎÓΟŌŌÒÖ0ΘÖOΘŎŎÖΘΟÓÒƠΟŐOƠOΘΘÓ0ÕΘŐƠÔÖΟΟ0OΘÖΘŎΟŐÕΘŐΟÓÓƠƠŐÒΟÖΘOŌƠΟΘŎ0ÒΟƠƠÔŎ00ŌÕÓÒ0ŌŎÒOƠŎÖŐÕÓƠ()
go ΟƠŐΟÔOƠÔΟΟÓΘÖOŐΟÕƠ0ΘŐÓÖÔŎÒ0OÔŐÓΟƠΘΘΘÔOÖŌÕŎÒŎΟÔÕÒŎÓ0OÖƠ0ÖÓOÒŐÓŐÔΘÓŌƠÕŎÕÒΟÖŌÒŎΟ0ÒÔÔÕ0ÕÓÒÓÒÔ0ŎO0Θ0ΘŐÖΟŎÖŌOΘÒΘ0ÕŐÖÓΟOÖÒÒÔÕOƠÒŎΘO0ÔÓÔ()
go ŎÓÕƠŌOŌΟÖÓ0ÒÓΘΟ0ƠŎÖΟŌŎÕÒÕÓŎÖÕƠÒŎŐƠÒƠ0ÒƠÖÕŎÒŐÓÔÓÒ0ΟÕŌÒŐOŐŌÖOÕƠƠÔŎÖ0ÒÓÒÕOÖÕƠÔÕOÔÖΟΘÕÔ0OÕƠÓOÓOÔÖÒ0ΟÖOÖÔΘŌOŌŎΘΘΘ0ŐŌOÔÒΟƠ0ŐÓÕŐÔŐΟÒŐÒΟ()
go ÔŌÕŌŐ0ŐÓÖΘŐÖΟŎΟƠŎOÖÖÕÒΘƠŐΟÓΟΘOÒƠΟ0ÒÒÖÖŌŎŌΘƠƠOΟÒƠΟΘOÓÕŌÕÒÕO0ŎOÔŐOÖŎÓÔOO0OÕŌOΘ0ÒƠƠÔŌÔÔÔŐΟ0ΘΘOŌOÒÕÒÕÕOÔÕŐŌΟÒŌÔ00ÖƠÔΟ0ΟÔŐΟƠÖÕΘ0ÖΟŌÖÖ()
go ÒŎ0ÖƠÔŌÔÖΟΘƠΟÒƠƠÖ0ŐÓÓΟOÓŎÖΟÔŌOOΟÔÒÓƠÓΟŎŎ0ÕƠŌÕÒÒÖ0OŌÕŐ0ƠÖΟƠÕÔÒÖŎÖÔÕÔŐOƠŌƠÒΘŌŎÖOŎΘÓÔÔΟŐŌΟŐƠÕΟÖŐ0ÔΘƠŌƠƠÒOOŎÔ0ŐΟÕÕŌÕÔÕÒŐÔŐÔÖÓƠÖΟOŌŎƠ()
ÒŎ0ÖƠÔŌÔÖΟΘƠΟÒƠƠÖ0ŐÓÓΟOÓŎÖΟÔŌOOΟÔÒÓƠÓΟŎŎ0ÕƠŌÕÒÒÖ0OŌÕŐ0ƠÖΟƠÕÔÒÖŎÖÔÕÔŐOƠŌƠÒΘŌŎÖOŎΘÓÔÔΟŐŌΟŐƠÕΟÖŐ0ÔΘƠŌƠƠÒOOŎÔ0ŐΟÕÕŌÕÔÕÒŐÔŐÔÖÓƠÖΟOŌŎƠ()
go ÒŎ0ÖƠÔŌÔÖΟΘƠΟÒƠƠÖ0ŐÓÓΟOÓŎÖΟÔŌOOΟÔÒÓƠÓΟŎŎ0ÕƠŌÕÒÒÖ0OŌÕŐ0ƠÖΟƠÕÔÒÖŎÖÔÕÔŐOƠŌƠÒΘŌŎÖOŎΘÓÔÔΟŐŌΟŐƠÕΟÖŐ0ÔΘƠŌƠƠÒOOŎÔ0ŐΟÕÕŌÕÔÕÒŐÔŐÔÖÓƠÖΟOŌŎƠ()
go ÔŌÕŌŐ0ŐÓÖΘŐÖΟŎΟƠŎOÖÖÕÒΘƠŐΟÓΟΘOÒƠΟ0ÒÒÖÖŌŎŌΘƠƠOΟÒƠΟΘOÓÕŌÕÒÕO0ŎOÔŐOÖŎÓÔOO0OÕŌOΘ0ÒƠƠÔŌÔÔÔŐΟ0ΘΘOŌOÒÕÒÕÕOÔÕŐŌΟÒŌÔ00ÖƠÔΟ0ΟÔŐΟƠÖÕΘ0ÖΟŌÖÖ()
go ƠΘÔŐŎŌÒŌÖŌΟŎOΟŐŎŎΘOƠÔ0ŐƠΘÒΟƠŐŌ0ŌΘΟΟÒŎÖÓ0ÒΟŐ0ƠΘŌÒÔÔŌƠŌÖOƠÓÔŌOÓ0OŐÕÒ0ŌÖÓŎÕŎÕÓÓ0ÒΘÖΟΘŎÕÓƠŐΟÕΟŌÔŎŌÓOÒÓ0Ŏ0ΟÕÔÒŌƠΘŌΟŎŐΘƠÔÓŐΘŎÔŌŌO0ƠÒÖÖ()
go ÒŎÒ0ΘŐŐÔƠΘÒŐŎΘÓÕÕÔŎÖ0ŌÖÖΟÖÓÔÓÒÖÖŌŐΘŌ0ÕÖŎÓΟŌŌÒÖ0ΘÖOΘŎŎÖΘΟÓÒƠΟŐOƠOΘΘÓ0ÕΘŐƠÔÖΟΟ0OΘÖΘŎΟŐÕΘŐΟÓÓƠƠŐÒΟÖΘOŌƠΟΘŎ0ÒΟƠƠÔŎ00ŌÕÓÒ0ŌŎÒOƠŎÖŐÕÓƠ()
go ΟƠŐΟÔOƠÔΟΟÓΘÖOŐΟÕƠ0ΘŐÓÖÔŎÒ0OÔŐÓΟƠΘΘΘÔOÖŌÕŎÒŎΟÔÕÒŎÓ0OÖƠ0ÖÓOÒŐÓŐÔΘÓŌƠÕŎÕÒΟÖŌÒŎΟ0ÒÔÔÕ0ÕÓÒÓÒÔ0ŎO0Θ0ΘŐÖΟŎÖŌOΘÒΘ0ÕŐÖÓΟOÖÒÒÔÕOƠÒŎΘO0ÔÓÔ()
go ÔŐŌŌÒÓ0OΟO0ÖΟŌ0ƠÒ0ÒŌÔOΟΟÖŎÕ0ΘÓΘOOΟΟΘ0Ơ0ΘŐÕ0ΟÔŐ0ŐŌÕ00ŌOΟΟÒÕÒÖŌÓOOÒÔΘÒΘÒÔŌŎΟÒŐŌÕÒÕŎÖÒÖÕΘ00OÖÓŌÓÖÖÓÔƠΘŌ0ÓÕ0ŌÖΘOÖ0ÓŐƠ0ΟÖ00ŌŐŐΘΘÕÒƠÕΟ()
go ŎÓÕƠŌOŌΟÖÓ0ÒÓΘΟ0ƠŎÖΟŌŎÕÒÕÓŎÖÕƠÒŎŐƠÒƠ0ÒƠÖÕŎÒŐÓÔÓÒ0ΟÕŌÒŐOŐŌÖOÕƠƠÔŎÖ0ÒÓÒÕOÖÕƠÔÕOÔÖΟΘÕÔ0OÕƠÓOÓOÔÖÒ0ΟÖOÖÔΘŌOŌŎΘΘΘ0ŐŌOÔÒΟƠ0ŐÓÕŐÔŐΟÒŐÒΟ()
go ÖƠŎŎΘΟÓÓÒÖÒ0ÓΘŎÖΘŌ0ŎOÕÔÓƠÔÔΘŌÔΘ0OÕOΟOÕƠŌÒŎ0O0ŎΘΟÒÕΟΟÓ0ŎΟÒΘÕΟÖŎÕÔOÕOΘΘŐÖŌŐΟÔΟÖΟ0OÔÓ0ΟŐOŌÕŌOŌOÕÒÒΘΟŎƠΟÒŌ0ƠΟÔOÒÓÒΟŐÕÒŌƠÔÖ0ŐÖÖÖÓΟƠŌÕ()
.....Another form of protection, is the dependency registration.
The idea behind it is to protect not only elf binaries, but also executable scripts, like python, bash, perl, php or anything with a shebang.
A common attack that is possible is a "man in the middle execution", for example, let the payload be a bash script, in the moment of execution /usr/bin/env bash is called to execute the content of the script.
If in the environment (or really with a symlink on /usr/bin/bash) we put something like:
#!/bin/sh
cp $0 /tmp/plaintext
/usr/bin/real-bash $@This will result in a transparent execution but a dumped plaintext in /tmp
(credits for the attack idea to mrnfrancesco (Francesco Marano) · GitHub )
So what can we do to ensure nothing like this happens?
We will register a binary/dependency that is necessary for the payload to run (for example /usr/bin/bash for a bash script, /usr/bin/python3 for a python script etc etc) and verify that the dependency in the target where we would execute the payload is valid or not.
Now there is a problem, how do we verify that it's a valid binary?
Using an Hash is too strict , for example /usr/bin/bash on fedora is different from the one on centos 8, but both are valid binaries.
Only the presence of it is not enough (like in the example attack before)
Some basic checks are made:
-
the file exists
-
file size is +/- 15% of the registered one
But something more is needed.
To address the problem it is possible to recycle a technique mostly used in the data-recovery territory: the byte frequency distribution study.
References:
The idea is to register the BFD of our dependency, then calculate the correlation index of Bravais-Pearson to see if the two dataset are linearly correlated.
Pearson's correlation coefficient is the covariance of the two variables divided by the product of their standard deviations.
If the index indicates strong correlation, we know that the two file are of the same type (binary)
A second step is to study the combined standard deviation of the two datasets (std deviation of the first and the second) and see if both values, the correlation and the combined std deviation are in certain ranges.
This ensures that the dependency we are finding is of the same type (a binary) using the Bravais-Pearson index and similar distribution from the combined std deviation.
Here an example of a valid dependency: /usr/bin/bash on fedora30 vs centos7
Blue and green the two BFD of the binaries, in green the combined std deviation.
This method is able to distinguish (and stop) the use of different (but functionally equal) binaries like zsh, dash or busybox
The last line of defense is also the encryption of the payload.
The decryption key of the payload is the sha512sum of the compiled launcher itself with the random garbage appended to it.
For example:
With offset 930000
Compiled launcher (800kb)
OFFSET1 (130kb)
Payload
OFFSET2 ( different one, 130kb)
The decryption key is
sha512sum(Launcher+OFFSET1)
This protects from file-based NOP attacks to remove some instructions from the launcher, and acts also as binary validation.
The payload is structured as above, OFFSET1 ends on the offset value, OFFSET2 is calculated as the byte reverse of the offset value.
The procedure will:
-
calculate the offset
-
take the launcher+OFFSET1, sha512sum it
-
remove OFFSET2 random bytes at the end
-
decrypt payload using the calculated sha512sum
So THE REAL DECRYPTION KEY IS BASED ON THE OFFSET ITSELF, all the obfuscation/anti-debug is to protect this information that is stored in an obfuscated string (that is not saved but computed at runtime) of random name and content.
As explained above, we will use a memory file descriptor to execute the binary without passing for the storage.
The binary will be executed using the Command library of Go, that uses the syscall exec under the hood, so no new shells are instantiated.
The process is launched and disowned.
Upon the termination of the launcher process, the /proc/PID/fd/mem of the process is deleted, so the only "copy" of the payload is in the process table of the detached process itself
This is the most secure approach as it deletes all instances of the plaintext payload, leaving it only in a highly private part of the RAM of the process itself, even gaining immunity to /proc dump attacks
This way even a script launcher with this approach is not possible to be retrieved:
Two routines are used to pipe the payload stderr and stdout to the launcher process.
Also piping is supported.
For IO heavy processes it is possible to insert in the Scan of the outputs an OB_CHECK like this:
// OB_CHECK
// async fetch stdout
go func() {
defer obWaitGroup.Done()
for obStdoutScan.Scan() {
// OB_CHECK
println(obStdoutScan.Text())
}
}()
// OB_CHECK
// async fetch stderr
go func() {
defer obWaitGroup.Done()
for obStderrScan.Scan() {
// OB_CHECK
println(obStderrScan.Text())
}
}()This will make an impact on performance (all check are executed for each IO on any standard output/error) but can give a layer of hardness to process hijacking or tracing
this is the least secure approach as while the launcher is running the /proc/PID/fd/mem of the process is still accessible and thus containing the plaintext payload ready to be stealed
This is particularly discouraged for srcipts payloads, as they are even easier to spot as the path is directly in the process
With binaries the approach is safer (only the binary name is in the process, thus making it hard to spot requiring a complete /proc dump)
