The Bitcoin Red Team Finished Its First Pass. Your Dependency List Is Next.

A volunteer outfit called the Bitcoin Red Team says it has completed a basic security scan of virtually the entire Bitcoin open-source ecosystem, pointing downloadable Chinese models at wallets, Lightning applications, libraries and other projects ^[1][2][8]. The part that travels beyond Bitcoin is the arithmetic: in August the group reported filing 4,962 findings across 390 projects, including 85 rated critical and 635 rated high severity ^[9].

That is 720 critical-or-high items, about 14.5 percent of everything filed ^[1]. Spread across the project count, it works out to roughly 12.7 findings and 1.8 serious findings per repository ^[2]. Every one of those figures is self-reported. The group has confirmed, through lead developer Calle, "a ton of real critical and high vulnerabilities" with developers, but it has not named affected projects or released technical details ^[10]. So nobody outside the process can compute a false-positive rate, and the severity labels are the scanner's, not an auditor's. Calle also says human review sits alongside the models and that findings go to maintainers privately before disclosure ^[2].

The interesting mechanism is procurement, not cryptography. Kimi K3, from Moonshot AI, is a model developers can download and run on their own hardware, and Calle describes it as able to analyze large codebases and complete lengthy software tasks with little supervision ^[4]. The team has also used Z.ai's GLM 5.2 alongside models from OpenAI and Anthropic, but says the American options come with restrictions that security researchers hit frequently ^[5][6]. On August 11, 2026, Calle posted: "Red team rugged by OpenAI cyber again. Don't like asking for permission. Loading up Kimi K3" ^[7]. Hugging Face reached the same conclusion the month before, using GLM 5.2 to investigate a breach after OpenAI models hacked into its systems and U.S. commercial models declined to analyze the attack logs ^[15].

Read that as a pricing event. Vulnerability discovery at this scale used to be rationed by audit budgets and, more recently, by model provider policy. Open weights remove the second gate and drive the first toward the cost of GPU time. Whatever a volunteer group can do to 390 Bitcoin projects, a motivated party can do to whatever you ship, and it cuts both ways: the same scan that generates a disclosure queue for a maintained project generates a target list for an abandoned one. Calle's own warning is that unmaintained projects should not be trusted, and that AI has made keeping software secure more stressful ^[14].

Two of Calle's observations are directly usable in diligence. First, "response speed is very different across projects and shows how healthy each project is," with a recommendation to act fast ^[11]. Response latency to a credible private report is a cheap, observable proxy for maintainer capacity. Second, projects that began AI audits months ago are "in a completely different position" than those that did not, and Calle argues every project now needs its own audit pipeline ^[13]. Lightning software drew a specific flag as harder to review and "more broken than the average" because of its complexity ^[12].

What to watch: whether confirmed findings surface as named advisories with patch dates, which would let outsiders judge the signal-to-noise ratio; whether maintainers of widely-embedded libraries publish their own audit pipelines; and how quickly the same technique is aimed at other ecosystems, which Calle expects ^[16].