- Galaxy Research confirms 1,596 BTC stolen and warns the total may reach 2,055 BTC.
- At least 15 separate attackers are now draining vulnerable wallets.
- Updating firmware does not fix a seed that was already created weakly.
A firmware flaw in Coldcard hardware wallets has let attackers reconstruct the private keys of thousands of Bitcoin holders and empty their wallets, with Galaxy Research counting confirmed thefts of 1,596 BTC from roughly 7,300 addresses and warning the figure could climb to 2,055 BTC, close to $130 million, if a suspected fourth wave is verified. The manufacturer, Canadian firm Coinkite, traced the problem to seed generation that quietly drew on weak randomness since March 2021. The first sweeps hit on July 30, and as of August 4 the theft is still running.
How thousands of “secure” seeds became guessable
Coldcard devices ship with a dedicated hardware true random number generator built into their STM32 chip, and that hardware-isolated key generation was a core part of the sales pitch. During seed creation, affected firmware never used it. A code path inside a library called libngu fell back to a software pseudo-random number generator named Yasmarang, and a preprocessor guard checked only whether a setting was defined rather than testing its value, so every build compiled against the fallback without raising a single warning.
The result was a private key drawn from a tiny, predictable pool. Instead of the 128 bits of entropy a standard seed should carry, Mk3 seeds effectively held around 40 bits, while the Mk4, Mk5 and Q recovered some randomness from their secure elements and landed near 72 bits. The Bitcoin Policy Institute put the gap in physical terms: a properly generated key sits somewhere among a set of possibilities so vast that finding it is like hunting one specific atom across two billion galaxies, whereas the affected Coldcard keyspace shrinks to roughly the number of atoms in a large virus, small enough for an ordinary computer to walk through in a few hours.
Affected
Affected
Safe
From one coordinated sweep to a swarm of attackers
The incident surfaced when about 594 BTC left roughly 500 single-signature wallets inside a 25-minute window early on July 30, hours before Coinkite published its first advisory. Galaxy Research then mapped the same morning more fully and put the first-day total at 1,082.65 BTC across 1,196 addresses, well above that opening estimate, before fresh waves pushed the count higher. What began as a single operator has splintered. Alex Thorn, who leads research at Galaxy, said on August 4 that the firmware is now being worked by numerous independent parties, estimating at least 15 different attackers racing to claim wallets before anyone else does.
Around 90% of the stolen coins have not moved, which investigators read as attackers waiting out the attention before attempting to cash out. Galaxy has passed roughly 600 suspected attacker addresses to U.S. law enforcement, exchanges and cyber-investigation groups, and set up a $5 million fund to support the response.
Why a firmware update alone will not save affected users
Coinkite has pushed emergency firmware for every affected model and says it destroyed its remaining vulnerable inventory, but patched firmware repairs future seeds, not past ones. A key generated with weak randomness stays weak even after it is restored onto a different brand of device, a point Trezor raised while reassuring its own customers. Anyone holding funds on an affected seed needs to generate a fresh seed on fixed firmware and move coins to it. Multisig setups offer no shelter if every device in the arrangement is a vulnerable Coldcard, since a quorum of secure devices is what actually protects the wallet. Holders who used a BIP-39 passphrase that was never typed into any computer face minimal risk from this specific flaw.
Coinkite also pushed back on reports that the current firmware permanently bricks devices. If the true-random-number-generator fault triggers, the company says the device fails closed with nothing written to flash, and a full power cycle restores it.
The $2 question hanging over the industry
Coinkite has said it must assume someone pointed AI at its open-source firmware to uncover the bug, and noted that its own AI review of the same code weeks earlier turned up nothing. A Reddit developer claimed Claude Code flagged the software-RNG fallback in about eight minutes. Dragonfly managing partner Haseeb Qureshi tested the idea more carefully and cautioned that the eight-minute result may have leaned on web search, while a separate offline run of GLM 5.2 reproduced the finding in roughly 20 minutes. His takeaway centered on price. That 20-minute run cost about $2, and Qureshi framed it through a “Cost of Discovery” metric measuring how cheaply a frontier model can surface a vulnerability. No evidence ties the actual thefts to AI, but the economics he described change who can afford to audit the code guarding large sums of money.
Whales are buying the fear while retail steps back
The fallout reached beyond the drained wallets. Santiment data shows Bitcoin wallets holding between 10 and 10,000 BTC added 19,610 coins since July 29, while addresses holding under 0.01 BTC reduced their balances by 0.55% over the same stretch. That cohort spans the whole market rather than any single Coldcard address, so the split reads as larger holders absorbing coins that nervous smaller holders are letting go as confidence in hardware self-custody takes a hit. Whether that confidence returns depends less on the loss tally than on what happens to the roughly 90% of stolen Bitcoin still sitting untouched on-chain, and on whether the swarm of attackers exhausts the vulnerable keyspace before affected users finish migrating.