The hardware wallet industry has long sold itself on a foundational promise: keep your private keys offline and your Bitcoin is safe. That promise took a serious hit this week after Coinkite, the Canadian company behind the widely used Coldcard hardware wallet, disclosed that approximately $38 million in Bitcoin had been drained from user wallets through a flaw in its key generation process — a flaw the company suspects was discovered by an attacker wielding artificial intelligence against its own open source firmware.
The scale of the loss is jarring on its own. Thirty-eight million dollars in Bitcoin, extracted through a vulnerability that existed quietly inside code that was, in principle, available for anyone to audit. That last detail is where the story turns genuinely uncomfortable for the broader self-custody ecosystem: the open source nature of Coldcard's firmware, long considered one of its core security virtues, may have become the precise mechanism by which this attack was enabled.
AI as an Offensive Security Tool
Coinkite's working theory is that the attacker didn't stumble onto the flaw through traditional security research. Instead, the company believes the perpetrator fed historical versions of its open source firmware into an AI system capable of systematically reviewing code at a speed and depth that would be impractical for a human researcher working alone. By scanning across the full version history of the codebase, such a system could, in theory, identify subtle regressions, logic errors, or cryptographic weaknesses that were introduced and later patched — but which remained exploitable in deployed devices still running older firmware versions.
This is a meaningful escalation in the threat model for open source hardware security. The community assumption has always been that openness creates a form of distributed protection — more eyes on the code, more chances to catch bugs before bad actors do. That calculus assumed a rough parity between defenders and attackers in terms of how efficiently code could be reviewed. Artificial intelligence dissolves that parity. A well-prompted large language model or specialized code analysis system can traverse years of firmware commits in minutes, pattern-matching for cryptographic anti-patterns or comparing function behavior across versions with a consistency no human team can match at scale.
The Key Generation Flaw
The specific nature of the vulnerability centered on key generation — among the most sensitive operations a hardware wallet performs. When a wallet generates private keys, the randomness and integrity of that process is everything. A flaw at that layer doesn't just expose one transaction or one address; it can compromise the fundamental security guarantee of every key ever generated under the affected firmware. The precise technical details of how the flaw operated have not been fully disclosed, which is consistent with responsible disclosure norms while investigations and potential recovery efforts remain active. But the category of vulnerability alone — key generation — signals that affected users may have had their wallets compromised from the moment of setup, not as a result of any subsequent operational mistake on their part.
This matters enormously for how victims understand their exposure. Users who followed every best practice — never connecting to the internet, using strong passphrases, keeping seed phrases physically secured — may still have lost funds because the weakness was embedded in the device's core cryptographic routine, below the level of any user behavior that could have compensated for it.
Open Source as Double-Edged Sword
Coinkite has built a strong reputation in the self-custody community precisely because it made its firmware open source, allowing security researchers and technically sophisticated users to verify its behavior. That transparency has historically been the right call. But this incident raises a question the hardware wallet industry will need to sit with: does open source firmware require a corresponding investment in continuous, AI-augmented defensive code review to remain genuinely protective in an era when offensive AI tools are accessible to well-resourced attackers?
The honest answer is probably yes. Publishing source code without systematically scanning version history for introduced-and-patched vulnerabilities — the kind that persist in deployed devices — creates an asymmetric window that sophisticated attackers can now exploit. Defenders know what the current firmware looks like. Attackers, armed with AI, may know what every version ever released looked like, and which ones are still running in the field.
What This Means for Self-Custody
The $38 million loss does not mean hardware wallets are broken as a category, and it would be a mistake to read this incident as an argument for keeping funds on centralized exchanges. The risk calculus of exchange custody versus self-custody remains complex. What this incident does demand is an urgent rethinking of security assumptions across the hardware wallet space — particularly the assumption that open source code is inherently safer without active, ongoing, AI-assisted defensive auditing of the full version history, not just the current release.
For Coldcard users specifically, the immediate priority is understanding which firmware versions are affected and whether their devices fall within the vulnerable range. Coinkite will face significant pressure to provide clear, actionable guidance on whether keys generated under affected firmware should be considered compromised and funds migrated to freshly generated wallets on verified clean firmware. The reputational stakes for the company are high, but the financial and personal stakes for affected users are higher. In the self-custody world, there is no fraud department to call.
Written by the editorial team — independent journalism powered by Bitcoin News.