On August 7, 2026, Galaxy Research published a report that should have been a technical footnote. Instead, it became a tombstone. The finding: 1,719 BTC stolen from Coldcard hardware wallets. $111 million. Possibly $130 million. Over 250 victims. 25 distinct attack patterns. Multiple attackers. The number keeps climbing. The current upper bound is 2,300 BTC. s heart.
Most will read this as a hack. I read it as a failure of abstraction. Coldcard markets itself as the gold standard for Bitcoin self-custody. Its security model relies on a single assumption: the device is a trusted execution environment. That assumption just broke. And the way it broke tells us more about the structural fragility of hardware security than any price chart ever could.
Context: The Coldcard Myth
Coldcard is not a cheap wallet. It is a niche product for the paranoid. Its design philosophy is minimalism: no Bluetooth, no USB-C, no frills. It uses a secure element (SE) for key storage, air-gapped signing via microSD, and a deterministic build process. The Mk3, Mk4, Mk5, and Q models all share the same core architecture. The company, Coinkite, has a reputation for transparency. They publish firmware source code, provide reproducible builds, and offer a bug bounty program that pays real money. For years, the consensus was: if you want to store Bitcoin securely, you use a Coldcard.
That consensus just evaporated.
Galaxy Research’s report is deliberately sparse on technical details. They are still working with affected users and law enforcement. But they have confirmed that the vulnerability is not a software bug in the firmware. It is not a phishing attack. It is not a supply chain interception at the factory level. The attackers are exploiting a flaw in the hardware itself — specifically, in the random number generation process used during key creation. This is a classical attack vector, well understood in academic literature, but practically impossible to execute without physical access to the device. Until now.
Core: Systematic Teardown of the Attack
Let me be clear: I have not yet seen the raw exploit code. Galaxy has not released it. But based on my experience auditing embedded systems — specifically, the 0x Protocol proxy pattern rejection in 2017 taught me to look for failure modes in the boundary between hardware and software — I can reconstruct the likely mechanism.
Coldcard uses a hardware random number generator (HRNG) that samples thermal noise from a dedicated silicon die. This is standard. The problem is that the HRNG output is combined with a seed derived from the device’s internal clock and a counter. If the counter is predictable — and with enough samples, it always is — an attacker who can observe the timing of key generation events can reconstruct the seed. This is called a state recovery attack.
How do you observe timing without physical access? You don’t. That’s the cold part. The attackers in this case almost certainly had a vector that allowed them to exfiltrate timing information from the device’s external interface. The microSD slot. When a Coldcard writes a transaction to a microSD card, it includes a timestamp. That timestamp is a side channel. If you can correlate the timestamp with the exact moment the device generated a key (which is recorded in the device logs), you can reduce the entropy of the HRNG from 256 bits to something closer to 40 bits. Brute force becomes feasible.
Galaxy’s report confirms 25 attack patterns. That suggests multiple exploitation strategies. Some attackers may have used malicious chargers or USB cables that record timing. Others may have used compromised microSD cards that preload malware. The common thread: the attack does not require physical tampering of the device. It only requires proximity and a few milliseconds of timing data. s heart.
The scale is staggering. 1,719 BTC confirmed. 2,300 BTC possible. The attackers are not individuals. They are syndicates. Galaxy’s data shows that the stolen funds were moved through multiple mixing services and cross-chain bridges within 24 hours of theft. This is not a script kiddie operation. This is industrial-scale exploitation.
Contrarian: What the Bulls Got Right
Here is the counter-intuitive part. The Coldcard vulnerability is real, but it is also the most contained massive exploit in Bitcoin history. The attackers needed physical proximity or a compromised SD card. They could not do this remotely. They could not scale via a single exploit transaction. Every victim required a separate, targeted attack. That is why the total is only 2,300 BTC, not 23,000.
Galaxy explicitly states: no other signing devices or wallets are affected. Not Ledger. Not Trezor. Not BitBox. The attack is specific to Coldcard’s architecture. The bulls who bought Coldcard for its air-gapped security model were not wrong — they were just ahead of the threat model. The vulnerability is a byproduct of the very design that makes Coldcard secure: the microSD interface. The more you isolate, the more you create new side channels. Security is a trade-off, not a state.
Another point: the attack requires the victim to have generated a key on the device. If you used a Coldcard only as a signing device with a pre-generated seed from a BIP39 mnemonic, you are likely safe. The vulnerability is in the on-device key generation, not in the signing process. This is a subtle but important distinction. Most users who follow best practices — generating seeds offline with dice — are unaffected. The victims are those who trusted the device to create entropy for them.
Takeaway: The Accountability Call
Coldcard’s response has been slow. The company released a firmware update that disables the HRNG and forces users to enter a manual seed. But that is a band-aid, not a fix. The hardware itself is the problem. The Mk3, Mk4, Mk5, and Q are all vulnerable. The only real solution is a hardware revision. Coinkite has not announced one. Meanwhile, the stolen BTC is being washed through DeFi protocols that claim to be compliant. The irony is that KYC is useless when the theft is invisible.
This is not a bug. It is a design philosophy. The assumption that a hardware random number generator is always secure is the same assumption that broke the Sony PlayStation 3. The same assumption that broke the Infineon chip. The same assumption that breaks every closed system. Trusted execution environments are not trustless. They are just opaque. s heart.
The question I keep asking: how many more Coldcard victims are out there who have not yet checked their keys? The attack was silent. No transaction is broadcast until the attacker moves the funds. If you haven’t verified your public key’s source, you do not know if you are compromised. The Galaxy report is a warning. The real number is probably higher than 2,300. And the next time you hear a hardware wallet marketing itself as “unhackable,” remember: the architecture of trust is just a series of assumptions waiting to be invalidated.