Hacking a Smart Home Device

How I reverse engineered an ESP32-based smart home device to gain remote control access and integrate it with Home Assistant.

Introduction

Recently, I've been slightly obsessed with connecting anything and everything in my house to Home Assistant. There's something so satisfying about having everything connected and automated in one application; I can finally forget every random mobile app for a different brand of smart product.

But there is one product I own that stubbornly doesn't connect to anything other than its own mobile app. It's a sleek air purifier that is unfortunately let down by its disappointing app.

In my spare time, I've been hacking together a project that removes the cloud dependency from smart devices just like this. So many modern products depend on an internet connection and cloud account for basic functions, and who knows what unnecessary data they collect or technical vulnerabilities they add to the home network?

I want to control this expensive air purifier just like the rest of my smart gadgets. And that marks the start of this challenging yet undoubtedly fun journey.

It's time to hack an air purifier! 😆

The Plan

If we're going to hack this device to be controlled by custom software, we're going to need to understand its current capabilities and plan a point of attack, requiring the least amount of work to achieve our goal.

The device already supports remote control with its own mobile app, which annoyingly requires a cloud account to use. By toggling my phone's Bluetooth, WiFi, and 5G, I was able to confirm that the app required an internet connection to control the device. Remote control was not possible locally via Bluetooth or WiFi.

This means the mobile app and device must be connected to a cloud server for the remote control to be possible. So, somewhere in that network, data between the device and its cloud server must be the fan speed and everything else the app controls.

So, that is our point of attack:

• If we can intercept the device's network traffic and change those values, we have control of the device.
• If we can emulate all of the server responses, we have control of the device without depending on an internet connection and its cloud server.

Mobile App Analysis

One of the first things I looked into was the remote control mobile app. This can be a quick way to gather some information, as Android apps can be relatively simple to pull apart.

Apps on Android are stored as a .apk file. With a quick search online, you can find a website to download a specific app's latest .apk. If you didn't know, the format of an .apk is technically a .zip file! you can simply extract them to browse the app's contents.

Android apps include compiled Java executables, usually named classes.dex. You can convert these to a .jar file with dex2jar and use jd-gui to browse the contents as reconstructed source code.

Locating the app MainActivity.class revealed that it is built with React Native!

package com.smartdeviceapp;

import com.facebook.react.ReactActivity;

public class MainActivity extends ReactActivity {
  protected String getMainComponentName() {
    return "SmartDeviceApp";
  }
}

For Android apps built with React Native, you can find the JavaScript bundle in assets/index.android.bundle.

A quick scan of the app's bundle revealed it uses a secure WebSocket connection:

self.ws = new WebSocket("wss://smartdeviceapi.---.com");

There isn't too much interest here in this Android app; as expected, it connects with their cloud server in order to remote control the smart device. It's worth a quick look due to the simplicity of getting some readable source code. We can always reference this bundle to see if any shared values or logic can be found there.

Network Inspection

Next up, it's time to have a look at the network traffic between the device and its cloud server; this is what we're trying to intercept and, ideally, emulate.

I use Pi-hole locally, which is a DNS server that blocks tracking and some ads, but it also has a useful feature to browse DNS queries by device. By navigating to the Tools > Network page and selecting the device's local network address, we can see it's querying the DNS server for the address of the cloud server's domain:

So now we know the cloud server's domain it's connecting to, we can use the Local DNS feature to send that network traffic to my local workstation (192.168.0.10) instead of their cloud server:

We can then use Wireshark to take a look at the traffic coming in from the smart device. We can do this by monitoring the workstation network interface with a filter of ip.addr == 192.168.0.61 (smart device address).

By doing this, I was able to see UDP packets being sent from the smart device to the workstation on the port 41014!

Packet Analysis

So, we know the smart device uses UDP to communicate with its cloud server. But right now, it's trying to communicate with my workstation and is expecting it to respond like its cloud server.

We can use a simple UDP proxy for our workstation to act as a relay between the smart device and its cloud server.

I used Cloudflare's DNS resolver (1.1.1.1) to look up the real IP address for their cloud server (because my Pi-hole DNS would have just resolved to my workstation's local IP address). Then I used node-udp-forwarder as a simple method to relay the traffic to their cloud server:

udpforwarder \
--destinationPort 41014 --destinationAddress X.X.X.X \
--protocol udp4 --port 41014

Looking at Wireshark again, we can see all the network traffic between the smart device and its cloud server!

When booting the device, it would send a packet to the server with data like this:

Hex View  00 01 02 03 04 05 06 07  08 09 0A 0B 0C 0D 0E 0F
 
00000000  55 00 31 02 01 23 45 67  89 AB CD EF FF 00 01 EF  U.1..#Eg........
00000010  1E 9C 2C C2 BE FD 0C 33  20 A5 8E D6 EF 4E D9 E3  ..,....3 ....N..
00000020  6B 95 00 8D 1D 11 92 E2  81 CA 4C BD 46 C9 CD 09  k.........L.F...
00000030  0E                                                .

The server would then respond with the following:

Hex View  00 01 02 03 04 05 06 07  08 09 0A 0B 0C 0D 0E 0F
 
00000000  55 00 2F 82 01 23 45 67  89 AB CD EF FF 37 34 9A  U./..#Eg.....74.
00000010  7E E6 59 7C 5D 0D AF 71  A0 5F FA 88 13 B0 BE 8D  ~.Y|]..q._......
00000020  ED A0 AB FA 47 ED 99 9A  06 B9 80 96 95 C0 96     ....G..........

All of the packets after this seemed to share a similar structure. They did not include any readable strings but were full of what appeared to be random bytes of data; this could be the Avalanche effect pointing toward encryption.

I searched around to see if this packet structure was an existing protocol. I read that DTLS is used by some smart devices and that it is based on UDP.

However, Wireshark does support the detection of DTLS packets but listed this packet as UDP, which means it couldn't determine a UDP-based protocol from the data. I double-checked with the DTLS specification, but that described a header format different from what we see in the packet, so we know DTLS isn't used here.

At this point, we hit a blocker; we don't understand how the data is formatted in these packets, which means we can't manipulate or emulate anything yet.

This would have been a lot easier if it used a well-documented protocol, but where's the fun in that?

Physical Disassembly

We know there are 2 applications that understand how to read this packet data: the smart device and its cloud server. And well, I don't have their cloud server handy, so it's time to take a look inside the smart device!

It was quite easy to disassemble with a few easily accessible screws. Inside was the main PCB containing the microcontroller, a port connecting to the fan, and a ribbon cable to the control panel on the front.

The main controller is labeled as an ESP32-WROOM-32D. This microcontroller is commonly used in smart devices and features WiFi and Bluetooth.

I stumbled across the ESP32-reversing GitHub repo, which contained a nice list of ESP32-related reverse engineering resources.

Serial Connection

The ESP32 contains a flash chip, which is where the firmware containing application logic is most likely stored.

The manufacturer of the ESP32 provides a utility called esptool to communicate with the ROM bootloader in the ESP32. With this tool, it's possible to read data from the flash, but first, we must establish a serial connection!

Referencing the ESP32 datasheet, we can find the pin layout diagram:

Here, we can see the TXD0(35) and RXD0(34) pins. We need to connect a wire to both of these pins and a ground pin for a serial connection.

The device PCB had a few pin holes, which are commonly connected to the pins for debugging and flashing; I was able to visually follow the traces from both of these serial pins to the holes! This allowed me to easily solder on breakout headers that I could temporarily plug jumper wires into. Otherwise, I would have likely carefully soldered directly to the chip pins.

With a multimeter set to continuity mode, I was able to locate which hole was ground by referencing the GND(38) pin on the ESP32.

Now, we need a port to handle this UART serial communication. I used my Flipper Zero, which has a handy USB-UART Bridge application under the GPIO category.

Using 3 jumper wires, I connected them together:

• Flipper Zero TX <--> RX ESP32
• Flipper Zero RX <--> TX ESP32
• Flipper Zero GND <--> GND ESP32

In Windows Device Manager, under the Ports (COM & LPT) category, I found my Flipper Zero UART device as COM7. Using Putty configured to a Serial connection on COM7 at 115200 speed, I was able to successfully connect to the Flipper Zero. While searching around, I saw this speed was often used for the ESP32, so I decided to go with it here.

When booting up the smart device, I noticed a bunch of log data from the serial output:

rst:0x1 (POWERON_RESET),boot:0x13 (SPI_FAST_FLASH_BOOT)
configsip: 0, SPIWP:0xee
clk_drv:0x00,q_drv:0x00,d_drv:0x00,cs0_drv:0x00,hd_drv:0x00,wp_drv:0x00
mode:DIO, clock div:2
load:0x3fff0030,len:4476
ho 0 tail 12 room 4
load:0x40078000,len:13512
ho 0 tail 12 room 4
load:0x40080400,len:3148
entry 0x400805f0
********************************
**    Starting SmartDevice    **
********************************
This is esp32 chip with 2 CPU core(s), WiFi/BT/BLE, silicon revision 1, 4MB external flash
Minimum free heap size: 280696 bytes
nvs_flash_init ret: 0
Running app from: factory
Mounting FAT filesystem
csize: 1
122 KiB total drive space.
0 KiB available.
FAT filesystem mounted
SERIAL GOOD
CapSense Init
Opening[rb]: /spiflash/serial
Serial Number: 0123456789abcdefff
Opening[rb]: /spiflash/dev_key.key
Device key ready
Base64 Public Key: **REDACTED**
Opening[rb]: /spiflash/SmartDevice-root-ca.crt
Opening[rb]: /spiflash/SmartDevice-signer-ca.crt
Addtimeout: 10000, id: 0
RELOAD FALSE
Opening[rb]: /spiflash/server_config
MP PARSE DONE
Server: smartdeviceep.---.com:41014

Dumping Flash

Awesome, now we have a working serial connection, we can focus on dumping the flash, hoping it contains information on how to read these packets!

To read the flash, we need to boot the ESP32 in a different mode, specifically what it calls the Download Boot mode. This is technically explained in the Strapping Pins section of the datasheet. But TL;DR, I held a jumper wire from a GND port on my Flipper Zero to the IO0(25) pin on the ESP32 while it boots.

Checking the serial output with Putty, we can see this successfully boots the smart device into the Download Boot mode:

rst:0x1 (POWERON_RESET),boot:0x3 (DOWNLOAD_BOOT(UART0/UART1/SDIO_REI_REO_V2))
waiting for download

Now we can close Putty and switch over to a Terminal to use esptool.

We're able to dump the entire 4MB of flash data from the ESP32 with the following command:

esptool -p COM7 -b 115200 read_flash 0 0x400000 flash.bin

I dumped the flash a couple of times to ensure I had a good read and backed them up in case we accidentally brick something because then we can flash back the dump.

Flash Analysis

We have the ESP32 flash dumped into a single binary file, and now we need to make sense of it. I found esp32knife to be the best utility for this.

It reads the flash file and extracts a bunch of useful information. It was also the only utility that successfully reformatted this dump into ELF format with correctly mapped virtual memory, but more on that later! Let's see what we can find:

python esp32knife.py --chip=esp32 load_from_file ./flash.bin

This logs out a lot of information and saves the output data to a ./parsed folder.

The first file of interest here is partitions.csv, this table maps areas of data in the flash:

# ESP-IDF Partition Table
# Name,   Type, SubType,  Offset,   Size, Flags
nvs,      data, nvs,      0x9000,   16K,
otadata,  data, ota,      0xd000,   8K,
phy_init, data, phy,      0xf000,   4K,
factory,  app,  factory,  0x10000,  768K,
ota_0,    app,  ota_0,    0xd0000,  768K,
ota_1,    app,  ota_1,    0x190000, 768K,
storage,  data, fat,      0x250000, 1M,

Device Storage

I was initially curious to see what data was in the nvs key-value storage partition.

The latest state of this data was extracted to part.0.nvs.cvs, and the only interesting data I could see was my WiFi SSID and password. But I also found the full historical changelog of values in part.0.nvs.txt and that revealed a couple of previously used WiFi credentials; what!? did someone use this thing before me?😆

Following that, it was time to look at the contents of the FAT storage partition. I found OSFMount to be a great Windows application for this; it mounts the filesystem image as a virtual disk and allows writing to it!

This revealed a few interesting files that we saw from the serial output earlier:

dev_info
dev_key.key
serial
server_config
SmartDevice-root-ca.crt
SmartDevice-signer-ca.crt
wifi_config

The dev_key.key file started with -----BEGIN EC PRIVATE KEY----- which is an Elliptic Curve private key; I used openssl to verify this with:

openssl ec -in dev_key.key -text -noout

And the two .crt files started with -----BEGIN CERTIFICATE----- which I also verified using openssl with:

openssl x509 -in ./SmartDevice-root-ca.crt -text -noout
openssl x509 -in ./SmartDevice-signer-ca.crt -text -noout

Having the certificates and device key stored on the device strongly indicates they are used to encrypt the UDP network packet data.

Initial Static Analysis

Now we've taken a look at the storage, it's time to look at the application which runs on the device.

We know it's running the factory partition, so I opened the part.3.factory file in the Ghidra CodeBrowser. Ghidra is a free and open-source suite of reverse engineering tools from the NSA; it's an alternative to the paid IDA Pro.

This file we're opening is the partition image direct from the flash; it's comprised of multiple segments of data, each getting mapped to different virtual memory regions on the ESP32. For example, data at offset 0x17CC4 in the partition image is actually mapped to 0x40080ce0 in the device's virtual memory, so although this file contains all of the application logic and data, Ghidra won't understand how to resolve any absolute memory references, at least for now. There will be more on this later!

The ESP32 microprocessor uses the Xtensa instruction set, and Ghidra has recently added support for this! When loading the image, you can select the language Tensilica Xtensa 32-bit little-endian. We can run the auto analysis; although it won't give us great results just yet, we can still look at any defined strings it is able to find.

String Theory

Text strings in a compiled application are a fast-track way of locating and understanding logic when reverse engineering; they can reveal a lot about the application.

Because this compiled file only contains bytecode instructions for the processor, there are no function names, data types, or parameters. It can initially seem like a giant blob of nonsense, but as soon as you a string reference like Failed to read wifi config file, you can start to piece together what the logic is doing. Reverse engineering compiled applications can be difficult, but it is certainly a rewarding challenge.

So, I had a look through the Defined Strings window in Ghidra to see what I could find, and noticed all of the strings we saw in the serial output, such as:

000031c4	"Serial Number: %s\r\n"
000031fc	"Device key ready\r"
00003228	"Base64 Public Key: %s\r\n"

In close proximity to these strings were some others of interest:

000030d0	"Message CRC error\r"
00003150	"Seed Error: %d\r\n"
000031c4	"Serial Number: %s\r\n"
000031fc	"Device key ready\r"
00003228	"Base64 Public Key: %s\r\n"
00003240	"Error reading root cert!!!!\r"
00003260	"Error reading signer cert!!!!\r"
00003280	"PRNG fail\r"
0000328c	"ECDH setup failed\r"
000032a0	"mbedtls_ecdh_gen_public failed\r"
000032c0	"mbedtls_mpi_read_binary failed\r"
000032e0	"Error copying server key to ECDH\r"
00003304	"mbedtls_ecdh_compute_shared failed: 0x%4.4X\r\n"
00003334	"Error accessing shared secret\r"
00003354	"####### MBED HKDF failed: -0x%4.4X ########\r\n"
00003384	"Sign failed\n  ! mbedtls_ecp_group_copy returned 0x%4.4X\n"
000033c0	"Sign failed\n  ! mbedtls_ecp_copy returned 0x%4.4X\n"
000033f4	"Sign failed: 0x%4.4X\r\n"
3f403d30	"Write ECC conn packet\r\n"

Ghidra Setup

Ok, it's about time we configure Ghidra to analyze this ESP32 application better.

First up, esp32knife supports reformatting the binary partition image for the application into an ELF format, which Ghidra can better understand. I had to make a small tweak for it to support the RTC_DATA segment, which I've pushed to my fork on GitHub: feat: add support for RTC_DATA image segment.

We can then import the more useful part.3.factory.elf instead of the part.3.factory binary partition image.

But when importing this time, we want to do a couple of things before running the auto analysis, so let's opt out of doing that for now.

Next, we can use the SVD-Loader-Ghidra script to import the peripheral structs and memory maps from the official esp32.svd file.

We can also use the built-in SymbolImportScript script to load labels for all ROM functions. I've published a file with all ROM function labels for the ESP32 ready for Ghidra here: ESP32_ROM_LABELS.txt. This will help us identify common ROM functions like printf.

Finally, we run the auto-analysis from the menu bar Analysis > Auto Analyze.

Let's see what that does to the strings we found earlier:

3f4031c4	"Serial Number: %s\r\n"
3f4031fc	"Device key ready\r"
3f403228	"Base64 Public Key: %s\r\n"

Firmware Modification

Up until this point, I have the PCB awkwardly positioned to keep the fan and control panel connected. So, I wanted to see if it would still function with them unplugged. Unfortunately, it did not; the serial logged the following:

I2C read reg fail1
No Cap device found!
REGuru Meditation Error: Core  0 panic'ed (IllegalInstruction). Exception was unhandled.
Memory dump at 0x400da020

Now we have Ghidra configured nicely, I took a look at the address mentioned in the log; it was assembly right next to a reference for the No Cap device found! string, and at the start of the function, it logs "CapSense Init\r". This must be for the control panel that uses capacitive sensing input!

I named this function in Ghidra to InitCapSense:

void InitCapSense()
{                       
  FUN_401483e0("CapSense Init\r");
  // ... CapSense logic
}

I then followed the references to this function back to another function that appeared to be starting as a task/service; I renamed this one StartCapSenseService:

void StartCapSenseService()
{
  _DAT_3ffb2e2c = FUN_40088410(1, 0, 3);
  FUN_4008905c(InitCapSense, &DAT_3f40243c, 0x800, 0, 10, 0, 0x7fffffff);
  return;
}

Again, I followed the function references and found the function that calls StartCapSenseService. Using Ghidra's Patch Instruction feature, I replaced the call instruction with a nop (no operation) instruction to remove the function call:

// Original
400d9a28  25 63 af    call8     FUN_4008905c

400d9a2b  65 31 00    call8     StartCapSenseService

400d9a2e  e5 37 00    call8     FUN_400d9dac


// Patched
400d9a28  25 63 af    call8     FUN_4008905c

400d9a2b  f0 20 00    nop

400d9a2e  e5 37 00    call8     FUN_400d9dac

We want to flash this change to the ESP32, so I replaced the bytes that were modified, not in this ELF file, but in the part.3.factory binary partition image, because that is in a raw format directly from the flash, so it will be easy to write back. I use a hex editor to find & replace the bytes:

2564af 653100 e53700 -> 2563af f02000 e53700

Then, I wrote this modified image to the ESP32 flash at the offset 0x10000, that is the offset from the partition table for the factory partition:

esptool -p COM7 -b 115200 write_flash 0x10000 ./patched.part.3.factory

But when trying to boot this, we get an error from the serial output:

E (983) esp_image: Checksum failed. Calculated 0xc7 read 0x43
E (987) boot: Factory app partition is not bootable

Alright, so there is a checksum. Luckily, the code inside esptool knows how to calculate this, so I threw together a quick little script to fix the checksums for an application partition image: feat: add image checksum repair script.

Now, we can use this to repair the checksums and flash the repaired image:

python esp32fix.py --chip=esp32 app_image ./patched.part.3.factory

esptool -p COM7 -b 115200 write_flash 0x10000 ./patched.part.3.factory.fixed

I tried booting the device without the control panel again; everything now works ok! We have successfully just modified the smart device's firmware!

Packet Header

Let's get back to focusing on the packets. We know the packets do not follow a well-known protocol, meaning we must figure out the structure ourselves.

I captured the packets from the device booting numerous times and compared them to each other. I noticed the first thirteen bytes were similar to other packets, while the rest of the packet seemed to be encrypted.

Here's the first packet received from the server between boots; you can see the data matches up until the offset 0x0D:

Hex View  00 01 02 03 04 05 06 07  08 09 0A 0B 0C 0D 0E 0F
 
00000000  55 00 2F 82 01 23 45 67  89 AB CD EF FF 37 34 9A  U./..#Eg.....74.
00000010  7E E6 59 7C 5D 0D AF 71  A0 5F FA 88 13 B0 BE 8D  ~.Y|]..q._......
00000020  ED A0 AB FA 47 ED 99 9A  06 B9 80 96 95 C0 96     ....G..........

Hex View  00 01 02 03 04 05 06 07  08 09 0A 0B 0C 0D 0E 0F
 
00000000  55 00 2F 82 01 23 45 67  89 AB CD EF FF 81 85 3F  U./..#Eg.......?
00000010  8A 10 F5 02 A5 F0 BD 28  73 C2 8C 05 71 6E E4 A3  .......(s...qn..
00000020  A6 36 FD 5C E0 D5 AC 3E  1A D5 C5 88 99 86 28     .6.\...>......(

It wasn't too difficult to figure out the first couple of values, then I noticed the remaining nine bytes matched the serial number from the device's serial output, and there we have the packet header format:

55 // magic byte to identity the protocol
00 31 // length of the packet in bytes
02 // message identifier
01 23 45 67 89 AB CD EF FF // device serial

The packets first sent and received had a slightly different format to those that followed; there were always the bytes 00 01 after the header in the client packet, and it was the only packet with the message ID of 0x02.

Comparing it to the other packets, I noticed a pattern with the message ID:

• 0x02 - First packet sent from smart device
• 0x82 - First packet received from cloud server
• 0x01 - All other packets sent from smart device
• 0x81 - All other packets received from cloud server

You can see the higher bits in this value represent if it's a client request (0x00) or a server response (0x80). And the lower bits are different between the first exchange (0x02) and all other packets (0x01).

Packet Checksum

We noticed a string in the application earlier that said "Message CRC error\r" which implied there is a CRC checksum in the packet. It would be helpful to know if there is a checksum in the data so it doesn't interfere with any decryption attempts.

I followed the references to this string, and a single function references it.

Let's take a look at the Decompiled code for that function:

// ...
iVar1 = FUN_4014b384(0, (char *)(uint)_DAT_3ffb2e40 + 0x3ffb2e42);
iVar2 = FUN_400ddfc0(&DAT_3ffb2e44, _DAT_3ffb2e40 - 2);
if (iVar1 == iVar2) {
  if (DAT_3ffb2e47 == '\x01') {
    FUN_400db5c4(0x3ffb2e48, _DAT_3ffb2e40 - 6);
  }
  else if (DAT_3ffb2e47 == '\x02') {
    FUN_401483e0(s_Connection_message_3f4030e4);
  }
  pcVar3 = (char *)0x0;
  _DAT_3ffb3644 = (char *)0x0;
}
else {
  FUN_401483e0(s_Message_CRC_error_3f4030d0);
  pcVar3 = (char *)0x0;
  _DAT_3ffb3644 = (char *)0x0;
}
// ...

We can see the s_Message_CRC_error label being used in the else block, so the if statement must verify the CRC data for a message.

This logic compares the results of 2 functions FUN_4014b384 and FUN_400ddfc0. If this is verifying the checksum of a packet, one must generate a checksum for the packet data, and the other must read the checksum value from the packet.

We could use the arguments to help us decide which is which, but let's take a look at both:

uint FUN_4014b384(int param_1, byte *param_2)
{
  uint uVar1;
  
  if (param_1 == 0) {
    uVar1 = (uint)*param_2 * 0x100 + (uint)param_2[1];
  }
  else {
    uVar1 = (uint)*param_2 + (uint)param_2[1] * 0x100;
  }
  return uVar1 & 0xffff;
}

uint FUN_400ddfc0(byte *param_1, uint param_2)
{
  uint uVar1;
  uint uVar2;
  byte *pbVar3;
  
  pbVar3 = param_1 + (param_2 & 0xffff);
  uVar1 = 0xffff;
  for (; pbVar3 != param_1; param_1 = param_1 + 1) {
    uVar1 = (uint)*param_1 << 8 ^ uVar1;
    uVar2 = uVar1 << 1;
    if ((short)uVar1 < 0) {
      uVar2 = uVar2 ^ 0x1021;
    }
    uVar1 = uVar2 & 0xffff;
  }
  return uVar1;
}

Now we know the last 2 bytes of every packet are a CRC-16 checksum and can exclude it from any decryption attempts!

Key Exchange

Earlier, we noticed mbedtls primitives labeled as ECDH and HKDF. So, what exactly are they?

ECDH (Elliptic Curve Diffie–Hellman Key Exchange) is a key agreement protocol that allows 2 parties (like the smart device and its cloud server), each having an elliptic-curve public–private key pair, to establish a shared secret over an insecure channel (UDP). I found a great explanation of this in more detail in "Practical Cryptography for Developers": ECDH Key Exchange.

Essentially, if the smart device and server generate an EC key pair and exchange their public keys, they can use the other's public key with their private key to compute a shared secret key. This shared secret key could be used to encrypt and decrypt the packets! And even though they exchange public keys over the insecure network, you still need one of the private keys in order to compute the shared key.

This is ideal for securing packets like this, and the first packet sent by the client is actually named the ECC conn packet in the logs:

UDP Connect: smartdeviceep.---.com
smartdeviceep.---.com = 192.168.0.10
UDP Socket created
UDP RX Thread Start
Write ECC conn packet

This is great progress; we know the first packet exchange is likely exchanging EC public keys to establish an ECDH key agreement to encrypt all the other packets.

If we ignore the packet header (13 bytes from the start) and checksum (2 bytes at the end), we can see the contents of the packets for this potential key exchange are both 32 bytes (256 bits), which would be a valid size for a public key. Even though the client's request has 00 01 at the start, we can assume this is some unimportant data descriptor as it doesn't change value between boots:

// Client request packet contents:

Hex View  00 01 02 03 04 05 06 07  08 09 0A 0B 0C 0D 0E 0F

00000000  00 01 D1 C2 B3 41 70 17  75 12 F7 69 25 17 50 4A  .....Ap.u..i%.PJ
00000010  C5 DD D4 98 06 FE 24 6B  96 FD 56 14 4A 70 7E 51  ......$k..V.Jp~Q
00000020  55 57                                            UW

// Server response packet contents:

Hex View  00 01 02 03 04 05 06 07  08 09 0A 0B 0C 0D 0E 0F
 
00000000  07 A8 02 73 52 42 1F 1F  C1 41 B4 E4 5B D9 A9 9A  ...sRB...A..[...
00000010  5A DD 0F 94 F1 AB 9E E8  86 C7 99 7E 08 68 52 C5  Z..........~.hR.

Ok, so what is the HKDF? That is HMAC-based key derivation. It can be used to convert shared secrets computed from Diffie–Hellman into key material suitable for use in encryption. Wow, that makes a lot of sense; it's most likely doing exactly that to derive a key to encrypt and decrypt the other packets.

Cryptography Analysis

To be able to decrypt these packets, we need to understand exactly how the key for encryption is generated. That includes any possible input data as well as configurable options.

It's safe to assume the ECDH and HKDF functions are used for the packet data, so focusing on the key generation process, I summarize the variables we need to understand:

• ECDH:
  • Public key
  • Private key
• HKDF
  • Hashing method
  • Output key size
  • Optional salt
  • Optional info

The smart device and its cloud server both exchange 256 bits of data during what we assume is the key exchange process. But remember, the smart device firmware also loads the following keys from storage:

• 256-bit device key pair (private & public)
• 256-bit cloud server "root" public key
• 256-bit cloud server "signer" public key

There are a lot of possibilities here, so I take another look at the application in Ghidra. By following the error strings, I located the function which generates this key! I steadily work my way through labeling functions and variables by comparing the assembly to the mbedtls source code. I was able to annotate and simplify it to the following pseudocode:

int GenerateNetworkKey(uchar *outputKey, uchar *outputRandomBytes)
{
  // Generate an ECDH key pair
  char privateKey1 [12];
  char publicKey1 [36];
  mbedtls_ecdh_gen_public(
    ecpGroup, 
    privateKey1, 
    publicKey1, 
    (char *)mbedtls_ctr_drbg_random, 
    drbgContext
  );

  // Overwrite generated private key?
  mbedtls_mpi_read_binary(privateKey1, (uchar *)(_DAT_3ffb3948 + 0x7c), 1);

  // Overwrite generated public key?
  mbedtls_ecp_copy(publicKey1, (char *)(_DAT_3ffb3948 + 0x88));

  // Load another public key?
  char publicKey2 [36];
  mbedtls_ecp_copy(publicKey2, (char *)(_DAT_3ffb38cc + 0x88));
  
  // Compute shared secret key using privateKey1 and publicKey 2
  char computedSharedSecret [100];
  uchar binarySharedSecret [35];
  mbedtls_ecdh_compute_shared(
    ecpGroup,
    computedSharedSecret,
    publicKey2,
    privateKey1,
    (char *)mbedtls_ctr_drbg_random,
    drbgContext
  );
  mbedtls_mpi_write_binary(computedSharedSecret, binarySharedSecret, 0x20);

  // Generate random bytes
  mbedtls_ctr_drbg_random(globalDrbgContext, outputRandomBytes, 0x20);

  // Derive key
  mbedtls_md_info_t *md = mbedtls_md_info_from_type(MBEDTLS_MD_SHA256);
  uchar* deviceSerialNumber = (uchar *)GetDeviceSerialNumber();
  mbedtls_hkdf(
    md, 
    binarySharedSecret, // salt
    0x20,
    outputRandomBytes, // input
    0x20,
    deviceSerialNumber, // info
    9,
    outputKey,
    0x10
  );
}

We now have a much better understanding of how the smart device generates the potential encryption key.

It's useful to keep in mind that their cloud server also has to generate this key, meaning it needs to have all the same input variables to the HKDF.

Knowing this, we can recap the three dynamic inputs to the HKDF function and understand how the server will also have them:

• salt - Shared secret: The server must have access to the same private and public keys used for the ECDH shared secret computation or use the public to our private and the private to our public.
• input - Random bytes: The server must have access to these randomly generated bytes on the smart device; either we send these bytes to the server, or technically, the server could recreate the pseudo RNG method used. However, the generated bytes have the size of 0x20 (32 bytes / 256 bits) which exactly matches the size of the data sent in the key exchange packet, so it's highly likely we're sending it there!
• info - Device serial number: We already know the device serial number is part of the packet header, so the server easily has access to this value.

Curious to know what the application did with these randomly generated bytes, I checked what the calling function did with them:

stack[0] = 0x00;
stack[1] = 0x01;
GenerateNetworkKey(&KeyOutput, stack[2]);
log(2, 2, "Write ECC conn packet\r\n");
SendPacket((int)param_1, 2, stack[0], 0x22);

Logging Key Data

Much progress has been made via static analysis, and the final value we need to calculate the decryption key is the shared secret.

At this point of reverse engineering, I hadn't reversed the functions as cleanly as shown in this blog post and wanted to try to dynamically obtain keys directly from the device.

Debugging via JTAG would be the sensible choice here. However, I didn't notice breakout points for these pins on the PCB, and I wanted to avoid soldering directly to the ESP32 pins, so I thought I'd challenge myself to patch the firmware to print it over serial!

The CapSense service is still disabled, so I thought I'd write a function over that logic to print out the shared secret key and call it right after it was computed!

So, planning in pseudocode, I'd want to add my function call to the GenerateNetworkKey function. Right after it has generated the key.:

int GenerateNetworkKey(uchar *outputKey, uchar *outputRandomBytes)
{
  // ... 
  
  // Add my function call:
  print_key(binarySharedSecret);
}

// Custom function saved over unused logic:
void print_key(char *key)
{
  for (int i = 0; i < 32; i++) {
    log("%2.2x", key[i]);
  }
}

While referring to the Xtensa instruction set architecture manual, I threw together some assembly like this:

// Original
400dbf2d  25 4b 6c    call8     GetDeviceSerialNumber

// Patched
400dbf2d  e5 ff fd    call8     print_key

// print_key:
400d9f2c  36 41 00    entry     a1, 0x20
400d9f3b  42 c2 20    addi      a4, a2, 0x20
400d9f3e  52 a0 02    movi      a5, 0x2
400d9f41  61 ea db    l32r      a6, PTR_s_%2.2x // "%2.2x"
400d9f44  d2 02 00    l8ui      a13, a2, 0x0
400d9f47  60 c6 20    mov       a12, a6
400d9f4a  50 b5 20    mov       a11, a5
400d9f4d  50 a5 20    mov       a10, a5
400d9f50  22 c2 01    addi      a2, a2, 0x1
400d9f53  25 ed 05    call8     log
400d9f56  27 94 ea    bne       a4, a2, LAB_400d9f44
400d9f59  22 a0 00    movi      a2, 0x0
400d9f5c  90 00 00    retw

I flashed the modified firmware, booted up the device, and checked the serial output:

Write ECC conn packet
e883eaed93c63d2c09cddebce6bb15a7f4cb5cedf00c1d882b8b292796254c9c

Success! We've printed out the shared secret key!

I rebooted the device numerous times to see if the key changed, and it remained the same. It is most likely computed using the keys in the firmware storage, but now we have the computed static value, we don't need to reverse the computation process.

Packet Decryption

Alright, we now understand the method to derive the decryption key and have all input values; it looks something like this:

const hkdfOutputKey = hkdf({
  method: 'SHA-256',
  salt: Buffer.from(
    'e883eaed93c63d2c09cddebce6bb15a7f4cb5cedf00c1d882b8b292796254c9c', 'hex'
  ),
  input: randomBytesFromDeviceKeyExchangePacket,
  info: deviceSerialNumber,
  outputKeySize: 0x10,
});

To be on the safe side, I wrote another firmware patch to print the key output from the HKDF call and tried recreating the key from captured packets. It works! That confirms we have correctly reverse-engineered the key creation function and are able to replicate the key creation logic in our own application.

But now we need to find which encryption algorithm is used. I refer back to the function which formats packets and found the call to the encryption function:

char randomBytes [16];

// Write device serial
memcpy(0x3ffb3ce0, deviceSerialNumber, 9);

// Generate and write random bytes
mbedtls_ctr_drbg_random(globalDrbgContext, randomBytes, 0x10)
memcpy(0x3ffb3ce9, randomBytes, 0x10);

// Write packet data
memcpy(0x3ffb3cf9, data, dataSize);

// Pad with random bytes
mbedtls_ctr_drbg_random(globalDrbgContext dataSize + 0x3ffb3cf9, paddingSize);

// Run encryption on the data + padding
FUN_400e2368(0x3ffb3cf9, dataSize + paddingSize, &HKDFOutputKey, randomBytes);

We know the key is 128 bits, with another 128 bits of additional random data.

I looked into the encryption function, which is very clearly crypto-related due to the looping of a bunch of bitwise operations, and noticed a reference to a static block of data.

This data started with 63 7C 77 7B F2 6B 6F C5, a search in the mbedtls source code revealed it is the AES Forward S-Box!

I decided to jump straight into attempting AES decryption on the captured packets and successfully decrypted a packet!! 🎉

Hex View  00 01 02 03 04 05 06 07  08 09 0A 0B 0C 0D 0E 0F
 
00000000  00 00 65 00 53 00 82 A4  74 79 70 65 AF 6D 69 72  ..e.S...type.mir
00000010  72 6F 72 5F 64 61 74 61  5F 67 65 74 A4 64 61 74  ror_data_get.dat
00000020  61 85 A9 74 69 6D 65 73  74 61 6D 70 CF 00 00 01  a..timestamp....
00000030  8D 18 05 31 FB A9 46 41  4E 5F 53 50 45 45 44 00  ...1..FAN_SPEED.
00000040  A5 42 4F 4F 53 54 C2 A7  46 49 4C 54 45 52 31 00  .BOOST..FILTER1.
00000050  A7 46 49 4C 54 45 52 32  00 07 07 07 07 07 07 07  .FILTER2........

The algorithm was AES-128-CBC and the additional random data was used as the IV (Initialization vector).

MITM Attack

We can now create an MITM (man in the middle) attack that does not require any firmware patching. This is because the private key of the device is now known, the key derivation logic has been reverse-engineered, and any required dynamic data is exposed over the insecure network.

If it correctly implemented ECDH, the smart device would have a unique private key that isn't exposed, and our easiest route of attack would be to generate our own server key pair and do any firmware modifications so the device accepts our custom public key.

But because of their custom protocol's design, we can write an MITM script that can intercept, decrypt, and potentially modify network communications without any modifications to the smart device. So, that's what we're going to do!

The main aim now is to decrypt and log as much data as possible; then, we can reference that to write a local server endpoint that entirely replaces their cloud server.

I hack together a quick Node.js script to do this:

const dns = require("dns");
const udp = require("dgram");
const crypto = require("crypto");
const hkdf = require("futoin-hkdf");
const fs = require("fs");

// Key Gen

const sharedSecretKey = Buffer.from(
  "e883eaed93c63d2c09cddebce6bb15a7f4cb5cedf00c1d882b8b292796254c9c",
  "hex"
);

function calculateAesKey(deviceSerialNumber, inputData) {
  return hkdf(inputData, 16, {
    salt: sharedSecretKey,
    info: deviceSerialNumber,
    hash: "SHA-256",
  });
}

// Packet Parsing

let latestAesKey = null;
let packetCounter = 0;
const proxyLogDir = path.join(__dirname, "decrypted-packets");

function decryptPacket(data, deviceSerial) {
  const IV = data.subarray(0xd, 0x1d);
  const encryptedBuffer = data.subarray(0x1d, data.length - 2);
  const decipher = crypto.createDecipheriv(
    "aes-128-cbc",
    latestAesKey,
    parsed.IV
  );
  decipher.setAutoPadding(false);
  return Buffer.concat([decipher.update(encryptedBuffer), decipher.final()]);
}

function logPacket(data) {
  const messageId = data.readUInt8(3);
  const deviceSerial = data.subarray(4, 4 + 9);

  if (messageId === 2) {
    // Key Exchange
    const randomlyGeneratedBytes = data.subarray(0xf, data.length - 2);
    latestAesKey = calculateAesKey(deviceSerial, randomlyGeneratedBytes);
  } else {
    // Encrypted Packets
    fs.writeFileSync(
      path.join(proxyLogDir, `packet-${id}.bin`),
      decryptPacket(data)
    );
  }
}

// Networking

dns.setServers(["1.1.1.1", "[2606:4700:4700::1111]"]);

const PORT = 41014;
const cloudIp = dns.resolve4("smartdeviceep.---.com")[0];
const cloud = udp.createSocket("udp4");
let latestClientIp = null;
let latestClientPort = null;

cloud.on("message", function (data, info) {
  logPacket(data);
  local.send(data, latestClientIp, latestClientPort);
});

const local = udp.createSocket("udp4");
local.bind(PORT);

local.on("message", function (data, info) {
  logPacket(data);
  latestClientIp = info.address;
  latestClientPort = info.port;
  cloud.send(data, PORT, cloudIp);
});

While capturing, I used their mobile app to remotely control the smart device so we could reference the logs and replicate the logic ourselves.

Data Exchange Format

We now have the decrypted packet data, but the data is still in a serialized binary format:

Hex View  00 01 02 03 04 05 06 07  08 09 0A 0B 0C 0D 0E 0F
 
00000000  01 00 64 00 29 00 82 A4  74 79 70 65 A7 63 6F 6E  ..d.)...type.con
00000010  6E 65 63 74 A8 66 69 72  6D 77 61 72 65 C4 10 00  nect.firmware...
00000020  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 83  ................

Looking back on this, I'm not sure why I didn't look for an existing solution that matches this serialized binary data format; I was expecting everything to be a custom solution at this point. But having a search now, this is just MessagePack, so I guess I just reverse-engineered and wrote a partial msgpack implementation 😆

Switching over to a popular implementation, we can see the data is easily unpacked into JSON:

const { unpack, pack } = require('msgpackr');

const packedData = Buffer.from(
  '82A474797065A7636F6E6E656374A86669726D77617265C41000000000000000000000000000000000', 
  'hex'
);

const unpackedData = unpack(packedData);

// unpackedData:
{
  type: 'connect',
  firmware: <Buffer 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00>
}

Network Log Analysis

In preparation for writing a custom local server for the smart device, let's take a look at the unpacked network logs we've captured:

🔑 Key Exchange Packet:

The smart device sends random bytes to the server to be used in the HKDF.

// Smart Device Request
D1C2B34170177512F7692517504AC5DDD49806FE246B96FD56144A707E515557

// Server Response
00000000000000000000000000000000

↙️ Get Device State:

The smart device fetches its initial state from the server when it boots.

// Smart Device Request
{ type: 'mirror_data_get' }

// Server Response
{
  type: 'mirror_data_get',
  data: {
    timestamp: 1705505010171n,
    FAN_SPEED: 0,
    BOOST: false,
    FILTER1: 0,
    FILTER2: 0
  }
}

🔗 On Connect:

When the smart device connects to the server, it sends its current firmware UUID. The server responds with the potential UUID for a firmware or config update that could be downloaded.

// Smart Device Request
{
  type: 'connect',
  firmware: <Buffer 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00>
}

// Server Response
{
  type: 'connect',
  server_time: 1706098993961n,
  firmware: <Buffer ab cd ef ab cd ef ab cd ef ab cd ef ab cd ef ab>,
  config: <Buffer 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00>,
  calibration: <Buffer 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00>,
  conditioning: <Buffer 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00>,
  server_address: 'smartdeviceep.---.com',
  server_port: 41014,
  rtc_sync: { ss: 13, mm: 23, hh: 12, DD: 24, MM: 1, YYYY: 2024, D: 3 }
}

⤵️ Server Updates Smart Device State:

When the server wants to update the smart device's state, it will send a packet like this.

// Server Request
{ 
  type: 'mirror_data',
  data: {
    FAN_SPEED: 1,
    BOOST: false
  }
}

⤴️ Smart Device Updates Server State:

The smart device sends its latest state to the server whenever it changes.

// Smart Device Request
{
  type: 'mirror_data',
  data: {
    timestamp: 1706105072142n,
    FAN_SPEED: 1,
    BOOST: false,
    FILTER1: 0,
    FILTER2: 0
  }
}

// Server Response
{ type: 'mirror_data' }

🛜 Keep Alive:

The smart device frequently sends a keep-alive packet to the server so the server can potentially use the open connection to send state updates.

// Smart Device Request
{
  type: 'keep_alive',
  stats: {
    rssi: -127n,
    rtt: 684,
    pkt_drop: 1,
    con_count: 1,
    boot_str: '',
    uptime: 100080
  }
}

// Server Response
{ type: 'keep_alive' }

MQTT Bridge

We're going to need a way to connect Home Assistant to our custom server, which handles the smart device networking. MQTT is ideal for this; it's a protocol designed for IoT messaging and can be easily configured within Home Assistant. For this, I set up the Mosquitto addon for Home Assistant, an open-source MQTT broker that connects everything together.

The connection chain will look like this:

Home Assistant <--> MQTT Broker <--> Custom Server <--> Smart Device.

The custom server logic in pseudocode would look something like this:

function HandleSmartDeviceRequest(req) {
  switch (req.type) {
    case 'mirror_data_get': {
      // Device wants state, send latest MQTT state or default fallback
      device.send({ fan_speed: mqtt.get('fan_speed') || 0 });
      return;
    }
    case 'mirror_data': {
      // Device state has changed, publish and retain in MQTT broker
      mqtt.publish('fan_speed', req.fan_speed, { retain: true });
      return;
    }
  }
}

function HandleMQTTMessage(topic, msg) {
  switch (topic) {
    case 'set_fan_speed': {
      // MQTT wants to change state, forward to device
      device.send({ fan_speed: msg.fan_speed });
      return;
    }
  }
}

The three main cases we are supporting here are:

• When the smart device boots and initially connects to the custom server, it requests the latest state; we can attempt to obtain this from the MQTT broker's retained value or fall back to a default state.
• When Home Assistant wants to update the state, it will send a command to the MQTT broker. We can subscribe to this command topic from the custom server and forward the request to the smart device.
• When the smart device's state changes for any reason, it sends the mirror_data packet to update the server state; we send this value to the MQTT broker to update the state and tell it to retain the data as the latest value.

I run this custom server alongside Mosquitto and Home Assistant on my small home automation server. Then configured my Pi-hole local DNS to resolve the cloud server's domain to my custom server.

Home Assistant Integration

The final step in this process is configuring Home Assistant to map the MQTT topics to a device type. For my air purifier, the closest integration was an MQTT Fan; in my configuration.yaml I added something like this:

mqtt:
  fan:
    - name: "Air Purifier"
      unique_id: "air_purifier.main"
      state_topic: "air_purifier/on/state"
      command_topic: "air_purifier/on/set"
      payload_on: "true"
      payload_off: "false"
      percentage_state_topic: "air_purifier/speed/state"
      percentage_command_topic: "air_purifier/speed/set"
      speed_range_min: 1
      speed_range_max: 4

Everything works! I've been running this for a couple of weeks now, and it has worked fine without any issues! I've even set up a little automation, so if my separate air monitor's PM2.5 or VOC level gets too high, it boosts the air purifier for a while!

Technical Recap

For better or worse, the engineers behind the service decided not to implement a standard protocol like DTLS. They created a custom solution which introduced some downsides to the system:

• We're not certain if each device has its own unique private key, but whether it does or not, both have downsides:
  • If all devices share the same firmware private key, the attacker needs to reverse engineer just a single device to MITM attack any other devices.
  • However, if every device has its own unique private key, the server must keep a data store mapping device serial numbers to the key of each device. So, In the case of any data loss, the server would entirely lose the ability to respond to any device communications; that is a scary thought for the business. Unless there is an insecure network fallback in place, which is equally alarming and time-consuming to develop
• Because the firmware contains a private key that is static, an attacker needs a single firmware dump to obtain the key and perform an MITM attack. Whereas, if an EC private key was instead generated at runtime, write access would be required in order to patch the server public key or application firmware, which could be protected by other means.

Also, the mobile app has a 1-star review on the app store. It makes me wonder if there is a correlation between the unexpectedly custom technical implementation and the abnormally poor end-user app experience. Building a custom system is far more than just the initial development; systems need support, and bugs need fixing.

Overall, it wasn't a bad implementation from a security perspective; you'd still need physical access to attack the device; there are pros and cons to everything and variables that aren't visible from our perspective.

The custom implementation increased the obscurity of network communication. However, Security through obscurity is simply a short-term win. While it may deter generic attacks on standard technical implementations. In the bigger picture, it's just an annoying yet passable hoop for an attacker to jump through.

I've had a few conversations recently about why engineers build from the ground up vs. using proven standards. And that's a very interesting topic; I'll save that for another post!

Conclusion

What a crazy journey that was!

I'd like to emphasize that the reverse-engineering process was not as smooth as it may seem from this post; I've done my best to format everything to be best read by you. But in reality, I was often in the dark, unsure if the next thing would work or not, and juggling many tasks and theories, iteratively making progress in multiple places to test my assumptions ASAP.

I tried some things that hit dead-ends and weren't worth dedicated sections in this post:

• I tried running the firmware in Espressif's fork of QEMU, patched out the CapSense service, and loaded virtual e-fuses to match the MAC address from the firmware, all to find out it doesn't support WiFi emulation. It was fun to see it booting virtually, though!
• I also tried flashing a different serial number, device key, and certificates to see if that affected anything before I got around to fully reversing the application logic. I didn't get much from this. Turns out this likely would have just affected the computed shared secret used for the HKDF salt, which we dumped anyway.

I've certainly sharpened a variety of skills from this project. I'm also proud I achieved my goal of adding this device to Home Assistant! The moment I managed to successfully decrypt the first packet was great; everything just clicked into place.

I'm still curious to explore creating an open-source project to de-cloud and debug smart home products; I've learned much more about the technical aspects of achieving that.

Thanks for reading! I hope you found some value in this post. I put a massive amount of effort into creating it, probably more than I did actually doing the project itself. It would be amazing to receive feedback on the format!

I'd also really appreciate it if you could help share the post.

You can drop a follow on Twitter to stay updated with what I'm doing.

If you'd like to show appreciation for my content financially, you can also buy me a coffee!

Take it easy 👋