The Machine Is Coming for Bitcoin's Keys
AI didn't kill Bitcoin. It may have killed the era when a catastrophic software flaw could remain hidden for five years.
Hundreds of people checked their supposedly cold, supposedly secure Bitcoin wallets and discovered the money was gone. Not transferred by mistake. Not lost because somebody clicked a phishing link. Not stolen because someone left a recovery phrase in Google Drive.
Gone because the hardware wallet itself had generated secrets that weren't nearly as secret as everyone believed.
The devices were Coldcards — among the most respected Bitcoin-only hardware wallets in the business. These are products bought specifically by people who understand that leaving cryptocurrency on an exchange is a little like leaving cash with a stranger who promises to watch it.
Cold storage was supposed to be the answer. Take the keys offline. Remove them from the internet. Put the authority to move the money inside a dedicated hardware device. Build a wall between the asset and the rest of the digital world.
It was a good wall. The problem was that somebody had accidentally built the key to the front door out of cardboard.
During an approximately 41-minute attack on July 30, around 1,200 wallets were reportedly drained in a coordinated sweep. Early estimates put the first wave near $70 million, with later analyses suggesting the total damage may have climbed substantially higher. The attacker appears to have prepared a list of vulnerable wallets, sorted them by balance, and started at the top.
This wasn't a guy in a basement manually trying passwords. This was industrialized theft.
And while there's no public proof that artificial intelligence discovered the vulnerability, the timing — and the nature of the flaw — should make every security professional, software company, and cryptocurrency holder pay very close attention.
Because the real story here isn't that AI killed crypto. It's that AI may be killing the shelf life of hidden mistakes.
THE SAFEST WALLET IN THE WORLD CAN'T PROTECT A BAD KEY
Cryptocurrency security ultimately rests on a fairly simple idea. Your private key is an enormous random number. The mathematical space from which that number can be selected is so incomprehensibly large that randomly guessing the correct one should be functionally impossible. Not difficult. Not expensive. Impossible for any practical purpose.
A hardware wallet can be disconnected from the internet, locked in a safe, and buried beneath Jimmy Hoffa. If the private key was generated correctly, the attacker still shouldn't be able to recreate it.
But randomness is doing a lot of work in that sentence.
According to Coinkite's technical disclosure, a software migration introduced in 2021 caused affected Coldcard devices to use the wrong random-number function during seed generation. Instead of relying entirely on the intended hardware source of cryptographic entropy, part of the process fell back to a software pseudorandom-number generator. On some earlier models, Coinkite now estimates that the resulting seeds may have provided only around 40 bits of effective security.
Forty bits sounds technical. Here's the nontechnical translation: the universe of possible keys became small enough to search.
The attacker didn't need to break into each Coldcard. They didn't need the wallets connected to the internet. They didn't need to compromise every owner's computer. They could reconstruct the possible keys somewhere else, derive the corresponding public Bitcoin addresses, and compare those addresses against the blockchain.
That last part is important. Bitcoin's ledger is public. Once you can generate a list of potentially vulnerable addresses, you can ask the blockchain which ones contain money. You can identify the live targets. You can rank them by balance. You can prepare the transactions ahead of time. Then you push the button.
One flaw. One search process. Hundreds or thousands of victims. Millions of dollars gone in minutes.
Sophisticated users may have done everything they were told to do. They bought a respected hardware wallet, kept the keys offline, and trusted that the device had created them correctly. That's what makes this different.
Air gaps protect good secrets. They can't rescue bad ones.
THIS WASN'T BITCOIN BEING HACKED
Let's get one thing out of the way before the tribal warfare starts. Bitcoin itself wasn't broken. The blockchain wasn't cracked. The consensus mechanism didn't fail. Nobody reversed the mathematics underlying the network. The attacker produced valid signatures using valid private keys, and Bitcoin processed the transactions exactly as designed.
The failure occurred further up the chain, inside the equipment entrusted to create and protect those keys.
That distinction matters technically. It doesn't make the event less serious.
Crypto people love to say, "Not your keys, not your coins." That remains true. But Coldcard just added a deeply uncomfortable second sentence: not your entropy, not necessarily your keys.
Self-custody isn't one technology. It's a chain of technologies and assumptions: the hardware, the firmware, the random-number generator, the software libraries, the compiler, the manufacturing process, the wallet application, the seed-generation ceremony, the backup method, the human being holding the thing.
Bitcoin can be mathematically sound while the machinery wrapped around it is catastrophically wrong. That's not an indictment of Bitcoin alone. Banking systems, authentication platforms, encrypted messaging applications, and government networks all depend on similar layers of software, hardware, and cryptographic implementation.
The difference is that when somebody compromises a bank account, there may be a fraud department, a transaction hold, an insurance process, or a regulator. When somebody reconstructs a Bitcoin private key, the network doesn't know there's been a theft. As far as Bitcoin is concerned, the owner moved the money. The key is the identity. The signature is the authority. There's nobody to call.
THE BUG SAT THERE FOR FIVE YEARS
This is where the AI question enters the room.
The vulnerability was reportedly introduced during a 2021 firmware migration. The code was publicly available. Engineers could inspect it. Security researchers could inspect it. Customers trusted it. The product accumulated years of use and reputation. The flaw remained.
Then, suddenly, somebody apparently understood it well enough to reconstruct vulnerable wallets, identify funded addresses, and coordinate a massive extraction operation.
There is currently no public evidence proving that an AI model found the bug. That must be said plainly. The attacker hasn't published a methodology. No forensic report has attributed the discovery to Kimi K3, Claude, ChatGPT, or any other model. Coincidence isn't evidence, and suspicious timing isn't attribution.
But come on. Look at the world around the event.
Frontier AI systems are now crawling software repositories, tracing dependencies, reading firmware histories, generating test harnesses, writing exploits, and working through problems that would once have required teams of specialists.
Anthropic has disclosed that its Claude Mythos research system developed a substantially improved attack against HAWK, an experimental post-quantum digital-signature candidate. For one proposed parameter set, the work required to attack it reportedly fell from roughly 2^64 operations to around 2^38. NIST subsequently confirmed that HAWK had been withdrawn from its post-quantum standardization process.
Mythos didn't break Bitcoin. It didn't break deployed AES. It didn't knock down the encryption protecting the internet. But it did contribute novel cryptanalytic work after years of expert human review.
That line matters. We aren't talking about an AI summarizing an academic paper about cryptography. We're talking about AI participating in the discovery of weaknesses that the people studying the system hadn't found. That's new. Or at least new enough, scalable enough, and cheap enough to change the entire security equation.
THE MACHINE DOESN'T NEED TO BE A GENIUS
I think this is where people misunderstand what AI changes. They imagine a singular, godlike model staring at Bitcoin's cryptography until it discovers a magic equation and steals the world's money. That's the Hollywood version.
The real version may be much more mundane — and much more dangerous.
An attacker doesn't need one omniscient machine. They need a collection of capable systems that can download old repositories, compare firmware versions, trace function calls, flag dangerous cryptographic patterns, identify duplicate function names, detect unexpected fallback behavior, generate tests, emulate hardware states, build candidate seeds, query public blockchains, rank potential targets, write transaction scripts, and repeat the process without getting tired.
The breakthrough may not be the model. It may be the scaffold around it. Give several imperfect agents tools, memory, a testing environment, and the ability to keep iterating, and suddenly they don't have to be Alan Turing. They only have to be persistent.
That's the part that should scare people. Human vulnerability research has always been constrained by time, specialization, and attention. There are millions of software projects and decades of old code. Even the world's best security researchers can inspect only a fraction of it. Machines don't have to choose one repository. They can chew through all of them.
PUBLIC CODE HAS BECOME A TIME MACHINE
Coinkite's CEO reportedly offered a brutal observation after the attack: if firmware is open source — or was ever public — you should assume attackers and defenders are already reading it with AI.
That doesn't mean open-source software suddenly became bad. Open source gives defenders the same visibility. It allows independent review, reproducible builds, public auditing, and faster collaboration. Closed-source code is hardly immune to reverse engineering, leaked repositories, compromised vendors, or malicious insiders.
But the old social bargain around public code has changed. For years, advocates could say that with enough eyes, all bugs become shallow. The uncomfortable reality was that most code never received enough qualified eyes. Now it may. Just not exclusively from the good guys.
Every forgotten GitHub repository, abandoned dependency, old firmware release, and archived code snapshot has become part of a permanent attack surface. AI doesn't only inspect what you're running today. It can inspect what you shipped five years ago. It can compare the old version with the patched version. It can infer what changed. It can identify customers who never migrated. It can reproduce the historical environment and test whether the same weakness still matters.
The internet never forgets. Now the machine can read.
DEFENSE GETS AI TOO — BUT OFFENSE ONLY NEEDS TO WIN ONCE
There is good news. The same systems finding vulnerabilities can help fix them. AI-assisted review can identify dangerous code before release, generate test cases, verify patches, inspect dependencies, and give smaller development teams access to capabilities they could never afford before.
Anthropic's HAWK discovery is technically a defensive success. The candidate was weakened before it became a deployed standard. That's exactly what cryptographic review is supposed to accomplish.
Coinkite has released fixed firmware for affected Coldcard models. But owners of potentially vulnerable wallets need to understand that a firmware update doesn't magically strengthen a seed that was generated years ago. They must generate an entirely new seed under corrected firmware and move the funds.
The bad news is that offense and defense aren't symmetrical. A defender has to find the bug, verify it, develop a patch, test the patch, publish the update, notify customers, and persuade those customers to migrate safely. An attacker only has to find the bug first.
In cryptocurrency, the attacker gets additional advantages: balances are public, transactions are global, settlement is rapid, transfers are irreversible, and victims may remain unaware until the money moves. One vulnerability can produce a complete target list. That makes crypto an almost perfect laboratory for machine-speed exploitation. Not because crypto is the only thing vulnerable. Because crypto offers the cleanest path from vulnerability to cash.
KIMI K3 AND THE UNCOMFORTABLE TIMING
This story has acquired another layer because the attack occurred shortly after the release of Moonshot AI's Kimi K3, a powerful Chinese open-weight model built for coding and agentic work. A senior White House official has publicly accused Moonshot of using large-scale distillation against Anthropic's Fable model to develop K3. The allegation is serious, but the supporting technical evidence hasn't been made public. Moonshot's precise acquisition of any Fable-derived cyber capabilities therefore remains unproven.
Could K3 or another model have helped find the Coldcard flaw? Absolutely. Do we know that it did? No. Anyone claiming certainty is getting ahead of the evidence.
But focusing too heavily on K3 risks missing the larger point. It doesn't matter whether this specific attacker used K3, Claude, ChatGPT, a private system, or no AI at all. The capability now exists across an expanding number of models and agent frameworks. Even if this particular robbery was performed entirely by humans, the next one may not be. The starting gun has already fired.
AI DIDN'T CREATE THE BUGS
This is the line people need to understand. AI didn't put the Coldcard vulnerability there. Humans did. The machine didn't weaken the random-number generator. It didn't approve the firmware migration. It didn't leave old devices running vulnerable code for years.
AI changes the discovery rate. For decades, software security has depended partly on a dirty little secret: there are more mistakes than people available to find them. We've been protected not merely by good engineering, but by attacker economics. Is this system worth studying? Does the researcher have the right expertise? Will finding the flaw take six months? Can the vulnerability be monetized? Is the target important enough?
AI crushes those costs. It can examine obscure systems that no human researcher would have chosen. It can cross-reference disciplines. It can revisit old assumptions. It can keep working after the human gets bored, goes home, or decides the project isn't worth the money.
The number of vulnerabilities may not suddenly increase. The number of vulnerabilities we discover will. That sounds like progress. Until the attacker gets there first.
WHAT CRYPTO HOLDERS SHOULD DO NOW
This isn't an argument for abandoning cryptocurrency. It's an argument for abandoning magical thinking.
"Hardware wallet" isn't a spell. "Cold storage" isn't a complete security architecture. "Open source" doesn't mean audited. "Military grade" doesn't mean invulnerable. "Air-gapped" doesn't mean the key was generated correctly.
Anyone holding meaningful cryptocurrency should be thinking in layers. Know exactly which device generated the seed and which firmware version it was running. Follow the manufacturer's security advisories. Understand whether an update fixes only future seed generation or protects existing wallets. Consider multisignature arrangements that prevent one defective device from becoming the single point of catastrophic failure. Don't generate every key in the same place, with the same vendor, and through the same process. Verify addresses on trusted hardware before signing. Protect recovery material physically and digitally. Treat experimental wallet software and browser extensions as attack surfaces.
And perhaps most importantly, stop assuming that a product's reputation proves the absence of undiscovered flaws. Coldcard had a strong reputation. That's why sophisticated people bought it. That's why the blast radius mattered.
THE AGE OF DORMANT VULNERABILITIES IS ENDING
The headline "AI Just Killed Crypto" is dramatic, clickable, and wrong. Bitcoin is still running. Cryptography still works. Cold storage still matters. Hardware wallets remain valuable.
But something important did die here. The comforting assumption that a catastrophic bug can sit quietly for five years because nobody has the time or ability to find it.
We're entering a world where every historical version of every important software project can be reviewed again — not by one researcher, but by thousands of tireless machine-assisted researchers operating on both sides of the line.
Some will report what they find. Some will patch it. Some will sell it. Some will wait until they've mapped every vulnerable target, sorted the balances, and prepared the transactions.
Then they'll press go.
AI didn't kill Bitcoin. It did something far more consequential. It put a clock on every mistake we've already made.
Sources:
https://www.coindesk.com/tech/2026/07/31/major-bitcoin-wallet-flaw-drains-594-btc-in-25-minute-sweep
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.
Comments