On July 30, an attacker emptied nearly 1,200 bitcoin wallets in 41 minutes, sweeping 1,082.65 bitcoin worth roughly $70 million per Galaxy Research’s on-chain mapping. No phishing, no malware, no physical access to any device. The wallets lived on air-gapped hardware; the keys were reconstructed offline by a machine that never touched them. By mid-August the high-confidence total had climbed to approximately 1,779 BTC (over $112–115 million at the prices when the coins were stolen) across more than 8,600 addresses, with three primary waves plus dozens of smaller attacker footprints. The cause was a flaw in the randomness behind the keys, sitting unnoticed in open-source firmware for more than five years. The lesson of the Coldcard exploit reaches well beyond one product: in a world where artificial intelligence can find and weaponize a dormant flaw at machine speed, any security model built on a single device, a single vendor, or a single moment of key generation is no longer defensible.
A five-year-old bug, and an industry that never looked
The technical failure was mundane, which is exactly what makes it damning. A March 2021 firmware change caused affected devices to skip their hardware random number generator during seed creation and fall back to predictable software-based key generation seeded by nonsecret chip data. Instead of the intended 128 bits of entropy, affected seeds carried as little as roughly 40 bits, a keyspace small enough to search offline. An attacker could regenerate candidate seeds by the billions, derive their addresses, and sweep whatever held a balance. Galaxy Research attributes 1,367 bitcoin across 4,585 addresses to a single operator, and nearly all of the stolen coin still sits unmoved.
Call this what it is: a gross failure of security review. Seed generation is the single most critical code path in a hardware wallet. It is the entire premise of the product. A one-line integration error in that path shipped in 2021 and survived five years of releases without anyone, vendor, auditor, or community, verifying that the hardware randomness the device was marketed on was actually in use. The theory that many eyes make all bugs shallow failed at precisely the spot where the eyes mattered most. And because a patch cannot repair a key that was born weak, every affected holder now bears the cost of a review that was never done. Singling out one company misses the point: almost no wallet vendor publishes its entropy architecture or submits that path to recurring adversarial audit, and there is no reason to believe this was the only such bug in circulation. 
AI changed the economics of exploitation
Weak-randomness incidents are not new. Android’s SecureRandom flaw drained wallets in 2013, and low-entropy brainwallets were picked clean for years. What is new is the cost curve. Bitcoin developers reviewing this incident noted publicly that the exploit was straightforward to reproduce with assistance from large language models. Work that once required a specialist team can now be substantially accelerated by tools available to anyone with a subscription. The on-chain wave pattern tells the same story: one operator’s July 30 sweep was followed within days by waves that researchers caution may be different actors, including roughly 443 bitcoin more on August 3. Once the keyspace was shown to be searchable, the exploit democratized in under a week.
That shift reprices every latent vulnerability in the ecosystem. A dormant bug was once protected partly by the scarcity of people capable of exploiting it. That protection is evaporating. The window between disclosure and industrial-scale exploitation is now measured in hours, and assumptions that were reasonable in 2021 deserve re-underwriting for 2026, starting with randomness, the invisible foundation beneath every key.
Redundancy is the answer, not retreat
One tempting response is resignation: if even respected hardware can fail, hand the keys to someone else entirely. Some analysts have already suggested the episode may push investors toward ETFs. But swapping one single point of failure for another solves nothing, and a bearer asset held entirely by a third party quietly stops being a bearer asset..
The other tempting response is to patch and carry on, trusting that open source caught the bug eventually. It did, but “eventually” was five years, and the draining began within minutes. Hoping the next dormant flaw is found by a researcher before a model finds it for an attacker is not a strategy.
The durable answer is the one every critical industry adopted long ago: engineered redundancy. Aviation does not certify a plane on a promise that engines never fail; it designs the aircraft to survive one failing. Bitcoin security should be held to the same standard. In practice that means no single source of entropy behind a key, no single device or codebase generating all of a wallet’s key material, and no single party whose one mistake can be fatal. Galaxy’s mapping shows the drained funds sat in single-signature wallets. Where an independent second key stood between the flaw and the funds, this exploit alone could not reach them. Decorrelation, meaning keys generated on independent hardware, firmware, and entropy sources, converts a catastrophic bug into a survivable one.
The standard to demand
The realistic objection is that redundancy adds complexity, which is why most people avoided multisignature setups. That was a defensible trade a decade ago. It is a poor one now that the cost of exploitation has collapsed while multi-key and multi-party custody tooling has matured across the industry. The burden should shift from the individual holder to the architecture.
The forward-looking question for every wallet maker, custodian, and serious holder is simple: if one component of your security failed silently today, would you find out from an audit or from the blockchain? Holders can act now by verifying how their keys were generated and adding genuine independence to their setup. Vendors can act by publishing entropy architectures, submitting to recurring independent audits, and making multi-source randomness the default rather than an expert feature. The machines are already searching. The only question is whether bitcoin’s security architecture is designed for the era in which they found something.
