Coldcard Hack: $100M+ Bitcoin Theft, AI Missed the Bug

A Five-Year-Old Bug, Hiding in Plain Sight

On July 30, 2026, attackers began draining Bitcoin from hundreds of wallets generated on Coldcard hardware devices, made by Canadian manufacturer Coinkite. The first wave moved fast: roughly 1,083 to 1,196 BTC pulled from about 1,196 addresses in just 41 minutes (The Hacker News; an early sub-window put it at 594 BTC, or roughly $38 million, drained in 25 minutes (CoinDesk). It was not an isolated event — a second and third wave followed over the following days, including roughly 208 BTC from 1,912 additional addresses flagged early the following Sunday (CoinDesk).

The root cause traces back five years. Coldcard MK3 firmware version 4.0.1, released in March 2021, contained a build-configuration error: the code checked whether a macro existed rather than whether it was enabled, which caused the device’s hardware random-number generator check to silently fail. Instead of using true hardware entropy, affected devices fell back to MicroPython’s deterministic “Yasmarang” software pseudorandom generator when creating wallet seeds (The Hacker News). The practical effect was a collapse in usable entropy from a full 128 bits down to roughly 40 bits — weak enough to be brute-forced with dedicated computing hardware, no exotic tooling required (1023Jack).

Blockchain analytics firm Galaxy Research, whose running tally has been cited by multiple outlets tracking the incident, put losses at approximately 1,367 BTC (~$89 million) from 4,585 addresses across three waves as of August 2 (CoinDesk). By August 3-4, a fourth wave pushed the tally to roughly 1,816 BTC (~$116 million) across more than 5,200 addresses, according to blockchain forensics firm TRM Labs, which cited Galaxy Research’s figures (TRM Labs). TechCrunch later reported losses exceeding $130 million as tracing continued (TechCrunch). TRM Labs is explicit that the number is not final: funds are still moving, and the firm “has not computationally confirmed that every identified address was generated with weak Coldcard entropy.” Treat every figure in this story, including this article’s, as a preliminary snapshot of an incident still unfolding.

Coinkite released emergency firmware on July 31, 2026 — MK3 to version 4.2.0 and later, MK4/MK5 to 5.6.0 and later, and the Q model to 1.5.0Q and later. The company’s guidance is unambiguous: anyone who generated a Coldcard seed between March 2021 and the patch should treat that seed as compromised and migrate funds to a newly generated one, regardless of current firmware version. One mitigation did hold up: “A seed built with at least 50 fair, independent, private dice rolls is not at risk from this bug alone,” Coinkite said, since dice-roll entropy never depended on the flawed hardware RNG path (Bitcoin Magazine).

The Deeper Lesson: Coinkite Tested Frontier AI on Its Own Code — and It Missed the Bug

The more solidly sourced, and arguably more unsettling, part of this story is not about the attacker at all — it’s about the vendor’s own admission. In a blog post titled “Adding to the Public Record,” published August 4, 2026, Coinkite disclosed that it had run frontier AI models against its own firmware, both before and after the breach, specifically looking for exactly this class of defect — and none of the models caught it (Coinkite).

The company’s own words: “We’ve run AI-assisted review against our critical codebases, including in the weeks before the exploit. It did not catch this vulnerability.” Then, after the breach, they went further: “Since the incident, we’ve also tested our code against frontier models, including Kimi K3, Claude Fable, and Codex 5.6. None of them caught it.”

That is a striking admission on its own terms. Coinkite wasn’t testing AI code review in the abstract — it retested with full knowledge that a specific, now-public vulnerability existed in its codebase, using multiple leading models including the same Kimi K3 that would later be linked (unconfirmed, see below) to online speculation about the attack. All of them missed it. The failure mode was structural: the flawed macro check “looked correct” at a glance, and the models tested apparently did not probe the build-configuration and submodule boundary where the actual defect lived (Coinkite). Coinkite drew a pointed lesson from that: “It’s a reason for us, and anyone else relying on AI tools, to be specific about what they currently catch and what they might not.”

That is the confirmed, on-the-record AI angle in this story — not that AI found or exploited the bug, but that multiple frontier models, run by the vendor itself and pointed directly at the vulnerable code, failed to find it either.

Did AI Find the Bug for the Attacker? That Claim Is Not Confirmed

Within hours of the first drain, a different theory spread rapidly across crypto and security social media: that Moonshot AI’s newly released open-weight model, Kimi K3, had been used by the attacker to scan Coldcard’s long-public firmware source and surface the vulnerability. The timing fed the theory — Kimi K3’s weights were released July 27, 2026, just two to three days before the attacks began (1023Jack).

This attacker-side claim has not been substantiated by any named investigator, security firm, or Coinkite itself. As one detailed technical postmortem put it: “No investigator has named any actor,” and “the Kimi K3 attribution is an unverified community claim” (1023Jack). The same analysis undercuts the premise on technical grounds: the entropy collapse this bug caused (128 bits down to roughly 40) is well within reach of conventional brute-force computing. “A 40-bit search needs no LLM; specialised hardware brute-forces it.” No AI sophistication was strictly required to exploit the flaw once it was known, and no named source has confirmed which tools, human or automated, the attacker(s) actually used to find it in the first place.

It is also important to be precise about what Coinkite’s own blog post does and does not say. Its AI-related statements, quoted above, are entirely about Coinkite’s own defensive review of its own code — the company never claims an attacker used AI, never names an attacker, and never identifies the tool or method used to find or exploit the bug. Some secondary coverage has blurred this distinction: one outlet’s headline asserted AI was “likely involved in the hack,” while its own body text hedges considerably, citing unnamed “industry experts” who “believe AI was used” without offering direct evidence (Bitcoin Magazine). The same article quotes Coinkite co-founder NVK making a general observation about the technology category, not a claim about this specific incident: “AI-assisted code review can now find latent bugs at a speed that is outpacing even the industry’s most seasoned experts.” That is a statement about AI code review’s growing capability — not a confirmation that it happened here, and not a claim that an attacker used it.

In short: it is confirmed that AI failed to catch this bug when Coinkite pointed it at their own code. It is not confirmed, and should not be reported as fact, that AI found or was used to exploit the bug on the attacker’s side.

A Related but Separate Story: A Post-Breach Volunteer Audit

In the days after the Coldcard news broke, a volunteer group of Bitcoin developers — led by pseudonymous Cashu protocol developer “Calle,” alongside Rob Hamilton, CEO of Bitcoin insurance company AnchorWatch — began running Kimi K3 as a large-scale, OpenSats-funded security audit across roughly 400 Bitcoin-ecosystem projects, spending about $10,000 per day on compute. Within 24 hours the effort surfaced roughly 5,000 potential issues, including 85 rated critical and 635 rated high severity, most subsequently verified by the affected project owners (Forbes). This was a direct, industry-wide reaction to the Coldcard shock — developers scrambling to find the next hidden bug before an attacker does — not evidence of how the Coldcard bug itself was found. No source reviewed connects this audit effort to the original discovery of the Coldcard flaw.

No CVE identifier has been published for this vulnerability as of this writing; Coinkite issued its own advisory and firmware updates directly rather than through the standard CVE/NVD disclosure process.

What This Means Beyond Crypto Wallets

Two separate, both-true lessons run through this incident, and conflating them does a disservice to either. First: open-source code being publicly auditable for years is not a guarantee that subtle build-configuration bugs get caught — Coldcard’s firmware was open source throughout the five-year exposure window, and no human reviewer, automated scanner, or (per Coinkite’s own account) AI tool caught this particular defect until after it was actively exploited. Second: the appeal of “an AI found and weaponized a years-old bug within days of its release” is a compelling narrative precisely because it plays into real, justified anxiety about AI-accelerated vulnerability research — but compelling is not the same as confirmed, and treating a viral theory as established fact is exactly the kind of unforced error that erodes trust in security reporting.

Frequently Asked Questions

Sources & Further Reading

• Coldcard Hardware Wallet Flaw Linked to $70 Million Bitcoin Theft in 41 Minutes — The Hacker News
• Coinkite Releases Fixed Firmware After Coldcard Bug; AI Likely Involved In The Breach — Bitcoin Magazine
• Adding to the Public Record on Our Ongoing Investigation — Coinkite Blog (primary source)
• Forty Bits: The Coldcard Hack And The Question Of Whether An AI Found It — 1023Jack
• The Largest Hardware Wallet Exploit of 2026: Inside the USD 116 Million Coldcard Hack — TRM Labs
• Bitcoin Cold-Wallet Attack Spreads to 4,500 Addresses as Losses Near $89 Million — CoinDesk
• Major Bitcoin Wallet Flaw Drains 594 BTC in 25-Minute Sweep — CoinDesk
• Hackers Steal Over $130M by Exploiting Bug in Offline Hardware Wallets — TechCrunch
• $38M in Bitcoin Drained by Coldcard Key Flaw Its Maker Thinks AI Found — Decrypt
• ‘Situation Is Extremely Bad’–Bitcoin Braced For More ‘Critical’ Exploits After $10,000 Price Per Day AI Warning — Forbes