Coldcard security flaw evaded AI code review

Coldcard's disclosure reveals a critical limitation in current AI-assisted code review for security-critical applications: the tools tested—including Kimi K3, Claude Fable, and Codex 5.6—all failed to catch a vulnerability that persisted across multiple firmware releases. The bug's location at a submodule boundary, rather than in core cryptographic logic, appears to have been the reason it remained invisible to both Coldcard's own AI-assisted review (conducted in the weeks before the exploit) and subsequent testing against frontier models. This suggests that AI code review tools may have systematic blind spots in detecting anomalies at integration points between separate code components, even when the flag check itself appeared superficially correct.

Coldcard's response—publishing a historical disclosure resource and recommending that other Bitcoin projects test AI tools specifically against build and submodule boundaries—indicates the company views this as an ecosystem-wide warning rather than an isolated incident. The statement that "many Bitcoin projects, including those that rely on open-source code, require immediate review" underscores the company's belief that similar vulnerabilities may be hiding in code elsewhere, undetected by the same AI tools that other teams rely on. This framing positions the flaw not as a failure of Coldcard alone, but as evidence of a gap in the current generation of AI security tooling that affects the broader Bitcoin hardware and software industry.