A practical supply chain attack on the Coldcard

TLDR: The Coldcard does a factory reset when an existing PIN is changed to an empty PIN , contrary to Coldcard’s claims that a factory reset is impossible. This can be used to distribute tampered devices without much effort. Coldcard has not patched the issue to date.

Preface

The Coldcard hardware wallet has a dual chip architecture, where some of the entropy used to generate keys and the PIN used by the user to authenticate is stored on a secure chip (an atecc608a), while the actual signing, transaction logic and i/o is done on a general purpose MCU (STM32L496RG).

I recently read that the Coldcard allows loading firmware that was not officially signed by the wallet’s developers. This is intended by them and even one of their advertised features. On their website coldcardwallet.com they state:

We have so much internal protection for the master secret, that we feel it’s safe to allow potentially hostile firmware onto this platform. If you don’t feel safe doing that, then it’s a choice you can make.

So what protects their users from malicious firmware updates in general? Firmware that is not signed by Coinkite will light up the red warning light the first time it is run. It also displays a “Danger” screen for 30 seconds to make sure users understand that something is wrong with the device:

Danger

These scenarios are also covered in their extensive documentation:

Obvious Hack-Attack

Idea: Find or steal a Coldcard. Load your trojan firmware onto it. Profit.

  • but you don’t know the main PIN (or else you’d have already stolen the funds)
  • so changing the firmware is not easy since it does little without the main PIN: you will have to crack the welded case, and do some difficult soldering.
  • regardless, once you change the firmware, red caution light will show
  • since you can’t set the genuine light without the PIN, and your trojan is signed with key zero, when the victim gets back, they will see the big “unauthorized firmware” warning, plus the red light and probably some scratches on the case, etc.
  • weak solution: “helpfully” power up the Coldcard for them, and say… “Here it is ready for your pin, sir. No idea why the light is red today.”
  • so we need to warn users to power up the Coldcard themselves every time they enter the PIN.

Finding a loophole

But what if for example a reseller of the Coldcard flashes his own firmware? In this case the user would not see the warning screen, or the red warning light, since it is only displayed on the first time the firmware is run. The main protection against this is that once the wallet has been setup, it cannot be reset again. This means that if the reseller would load his own firmware, the user would notice, because he would receive an already setup device locked with the reseller’s PIN. As quoted from the their faq:

Is there a factory reset?

There isn’t a factory reset due to the secure element. Most of the fields in that chip cannot be quickly reset. However, you can clear the wallet seeds and remove secondary PIN code individually. It’s a lot of typing and all corresponding PIN codes must be already known to you.

They also go on to say:

if new device is intercepted from our factory (ie. without a main pin set), new code cannot be loaded until the PIN is set, and there is no way to clear main PIN.

In the following I will show that both these claims are not true. There is indeed a simple way to clear the main PIN in the firmware, which in turn does a factory reset. This allows for a supply chain attack (or as they call it interception from the factory).

The exploit

I then checked in the Coldcard source code if this claim actually holds up, attempting to find a way with my own, changed firmware to put the device back in what I would call “Blank - fresh off the factory” mode - meaning in all practicality “factory resetting” it, or putting it back in the same state as when the device arrived in the mail. This would allow an attacker to load tampered firmware and distribute the tampered device in “mint” condition to his victim.

Its bootloader code is not that straight forward. It consists of multiple state machines written in C, that produce state values that are read out by some micropython code. While the C (or bulk) part of the bootloader is in stm32/bootloader, the micropython (or user facing) code is in shared.

When booting up, the wallet first goes through a setup function that checks the PIN monotonic counter and initializes communication with the secure chip. It then checks if the main PIN is blank (filled with zeroes) in stm32/bootloader/pins.c:

575 // need to know if we are blank/unused device
576 if(pin_is_blank(KEYNUM_main_pin)) {
577     args->state_flags |= PA_SUCCESSFUL | PA_IS_BLANK;

If it is blank, the PA_IS_BLANK state is set and is read out by the micropython code. In the relevant case, this is in shared/actions.py:

681 elif pa.is_blank():
682     # let them play a little before picking a PIN first time
683     m = MenuSystem(VirginSystem, should_cont=lambda: pa.is_blank()

As we can see, if the PIN is indeed blank, the system is set into a “virgin”, or fresh off the factory state. The user is then displayed the menu to set the PIN, or show the Bag Number. This means that if an attacker loads tampered firmware and then lets that firmware change the PIN to zero, the device goes back to this “blank” state. The PIN relevant PIN change code is in shared/actions.py:

1281    if pin is None:
1282        return await ux_aborted()
1283
1284    is_clear = (pin == '999999-999999')

In the tampered firmware I tried to set the PIN under this line to pin = ''. Indeed, it did not check if the changed PIN provided is empty, and will write it as the new PIN to the secure chip memory slot.

On next startup it will then the go through the normal setup procedure - as if for a mint condition device. If we just change the PIN it will however not show the accept terms screen that is usually shown on a fresh device. The value keeping track of whether the terms have been accepted before or not can be manipulated from our malicious firmware as well. Meaning if we add settings.set('terms_ok', 0) to the tampered firmware, the terms will display again on next startup.

At this point a reseller could then distribute the Coldcard to his customers, without a trace that it has been tampered with. They later state in their faq that:

All legitimate resellers should be providing the Coldcard unused and still in it’s original tamper-evident bag. As part of the first-use sequence, you will verify the bag number matches the factory bag number.

Since the device is the same, and the bag number is not changed in the device’s memory, the user will still be able to verify the device with the same bag number. This means the attack also breaks a key feature of their supply chain security. To summarize:

  • The attack has virtually no costs, scales easily and is quick to execute.
  • It reduces the supply chain security to only that of the plastic sleeve.
  • There are no limits to what this tampered firmware could do other than the wallet’s hardware constraints.
  • Compromising a future user’s funds would be as easy as setting the BIP32 root private key to that of the attacker.

Mitigation

There don’t seem to be any easy mitigations against this attack on current Coldcard devices, since it would require changes to the bootloader, which is locked-down and therefore cannot be updated easily. A patch can only be deployed to new devices shipped with a patched bootloader. As a future mitigation for new devices, they could not allow its bootloader to run arbitrary firmware. From what I gathered so far, few users load their own, self-compiled firmware. Another potential mitigation might be to have a different flag, or even a slot in the secure chip to track if the device has been setup before. It should also check that the changed pin fulfills some minimum requirements, especially being non-zero.

So how can a current user protect herself from this exploit? In theory, even when the user attempts to upgrade the firmware with the genuine, developer signed-off binary, the malicious firmware could play nice and update it by putting it into a flash region that is not checked by the bootloader. The bootloader only checks regions as defined in the header of the firmware (which sits at the start of the firmware chunk in memory). Writing such a malicious firmware would be involved though (I screwed up on first attempt and now it’s a brick), especially because of the 1 MB flash limit set in place by the STM32L496RG mcu and the code execution protections on the chip. This means that as a band aid solution all users should update their firmware first before setting up the device.

Normally I’m not interested in the physical security of hardware wallets, but since they argued to me that enough physical security is provided as is, I wanted to highlight that I strongly contest this claim. Not only is the factory bag number left intact by the exploit, but the plastic sleeve itself is easily compromised with minimal traces and household tools by cutting it open at the bottom with a sharp knife and resealing it again with heat in a similar way, or if laminating equipment is available in the same way, like the sides of the sleeve.

Coldcard’s response

After initially acknowledging the issue, no fix was provided for it and no bounty was paid out. Since they have no intention of fixing the issue, I am now publishing this report. I did ask for prior permission to post about the issue, which they gave me as long as I mentioned their PIN code developer documentation (also linked and quoted further above): https://github.com/Coldcard/firmware/blob/master/docs/pin-entry.md#how-to-develop-professional-code-on-coldcard

Update 2020-05-17

Coldcard has now published a blog post addressing the issue. They claim that the attack I described was part of their initial threat model. This is absurd, since the main claim that the PIN cannot be changed (as quoted further above) was changed the day they published the blog post in this commit. They also present updating the firmware as a golden bullet mitigation against the attack. Again, while I think updating the firmware is a decent protection, this alone is not enough. Users also need to a use a feature that was introduced in that update, to be sure that the update was installed in the first place. I also still believe that a secondary, thin bootloader acting as a permanent backdoor can be flashed on the device, however I lack the motivation and tools to attempt this.

Written on May 12, 2020