Someone likely used AI to drain almost 600 BTC, worth $38 million, from roughly 500 dormant wallets yesterday as part of a seed phrase exploit targeting Coldcard hardware wallets.
The attack took just 25 minutes to move the BTC from 500 single-signature addresses into a single address, and reports suggest the exploit will likely continue.
Coindesk reports that the affected BTC was dated between 2021 and 2026, and much had remained dormant for years. Of the 594 coins stolen, 562 remain in the same address at the time of writing.
Coldcard maker, Coinkite, confirmed hours after the exploit that seed generation within its Mk3 wallet, and its subsequently updated versions beyond March 2021 (version 4.0.1), may not have been random at all.
Coldcard initially claimed that its Mk3 devices were at risk, and that the Mk4, Q, and Mk5 are “not affected based on our early analysis.”
Read more: Credit default swaps forecast AI bankruptcies
Block, formerly known as Square, found different results in its published analysis while one of its team members, Max Guise, found flaws between Mk2 and Mk5 Coldcard models.
The payments company traced the bug to a mis-written compile-time check. The newer devices, Block found, carry a smaller, but real, version of the same flaw.
Coinkite believes AI was used to discover exploit
Coinkite’s recent analysis deduced that, because Coldcard’s source code is open and public, someone likely used AI to exploit it.
It said that a few weeks before the attack, it couldn’t spot the bug even with Coinkite’s use of “the best available AI models.”
It added, “Both attackers and defenders have the same AI tools, but today it did not help us, and only helped the bad guys.”
Pseudonymous owners of Bitcoin.org website, Cobra, also expressed that they have “very bad feeling AI was involved,” and noted, “For whatever reason some addresses are only being partially drained despite the private key being compromised. Strange.”
Read more: Apple threatens Sparrow bitcoin wallet dev with App Store termination
Crypto developer Stephen DeLorme claims he was able to use AI model Claude Opus 5 to sniff out the Coldcard vulnerability after cloning the firmware’s repository.
“All our software is insecure, and we’re painfully figuring that out in realtime with AI agents,” DeLorme said.
The technicalities behind the Coldcard BTC theft
BTC wallets need genuinely random numbers to generate an unguessable private key. Coldcard’s firmware was supposed to pull that randomness from a hardware generator built into its STM32 chip.
According to Block, a codebase check tested only whether a macro called MICROPY_HW_ENABLE_RNG was defined, not what value it held.
Coinkite’s software build set that macro to zero on purpose. Because the character was set to zero, and not a variable symbol, the check was flawed.
Despite this, the flawed check passed anyway during software operations. Firmware fell back to Yasmarang, a MicroPython pseudo-random number generator never meant for real-world cryptographic protection.
Read more: The number of BTC wallets holding more than 0.1 BTC hasn’t grown in two years
Bitcoin Core developer Gregory Sanders reproduced the attack using setup button-press counts, and confirmed its impact on Mk3 and Mk2 models. His own response to his findings was, “Sorry, this is the time to panic.”
Sanders first wrote, “confirmed. Mk2/3 vuln, I don’t think mk4 is but can’t be certain,” before following up an hour later with “mk4 is probably not much better.”
Got a tip? Send us an email securely via Protos Leaks. For more informed news and investigations, follow us on X, Bluesky, and Google News, or subscribe to our YouTube channel.
The post Coldcard attack: 25 minutes, 500 wallets, $38M in BTC gone appeared first on Protos.
