AINEWS
The 4,900 flaws, 85 of them critical, hit 390 open source Bitcoin projects and were found by Chinese AI because U.S. acceptable use policy blocks the prompts.
A volunteer collective called the Bitcoin Red Team reported 4,900 vulnerability findings across 390 open-source Bitcoin projects during August, according to the group's pseudonymous lead, Calle, and a summary at ChinaTechNews. Of the reported flaws, 85 were classified as critical and 635 as high severity. The targets are not named publicly; they span the categories the team audits every month: wallets, cryptographic libraries, Lightning software, and the supporting infrastructure that lets the network run.
The audit was not a U.S. operation. The models that did most of the scanning are Chinese: Moonshot AI's Kimi K3 and Z.ai's GLM 5.2, with OpenAI's GPT Sol and Anthropic's Claude Fable and Opus picking up the rest. The reason, Calle said in an August update on X, is that U.S. vendors' acceptable-use policies block the dual-use prompts the audit depends on. "We're experiencing a massive collision between decades of human open source slop against 2 weeks of Kimi K3," Calle wrote.
That single line captures the inversion this audit exposes. The Bitcoin stack is open source. Anyone with the right tool can read it, and anyone with the right tool can also find flaws in it. The same model class that is doing the auditing here is the one a well-funded attacker would reach for. Defenders and adversaries are running the same audits at the same latency, and the side that finishes first gets to decide whether the flaw is patched or exploited.
The team is small and the spend is modest. Rob Hamilton co-runs the project on X. Calle, who uses a handle rather than a real name, is the technical lead. By Calle's own count, posted to X, the group has now scanned about 150 Bitcoin repositories, filed more than a dozen private vulnerability disclosures with maintainers, and burned roughly $20,000 on AI services. The daily run rate is around $10,000. The ChinaTechNews summary cites a different August snapshot: 4,900 findings across 390 projects, with 85 critical and 635 high-severity issues. Both figures are self-reported, not third-party-verified, and they should be read as the same project at two different points in the month rather than a single reconciled total.
The license-to-operate gap is the load-bearing fact. U.S. model vendors treat offensive-security prompts as a misuse category. Anthropic and OpenAI's acceptable-use policies restrict code that probes for exploitable flaws in third-party software without permission. Chinese vendors, as a matter of public policy, are not enforcing the same line. The result is not that Chinese models are better at the work; it is that they are available for it. The Bitcoin Red Team's choice is a tooling-fit decision, not a quality judgment.
Hugging Face used GLM 5.2 to investigate a post-intrusion breach, and earlier in the year Anthropic's Claude Opus 4.8 helped researchers find a four-year-old Zcash flaw that could have minted unlimited ZEC, as Decrypt reported. Coinkite, the maker of the Coldcard hardware wallet, has said it believes attackers used AI to find the recent Coldcard vulnerability, suggesting the same tooling the Bitcoin Red Team uses is what well-resourced adversaries reach for.
The harder question is what the maintainer side does next. Most of the projects being audited are maintained by a handful of volunteers, and a private disclosure from a red team is a triage problem, not a fix. A dozen new bugs in a month, in code that is the only thing standing between a user's keys and a thief, is a workload those projects do not currently have the staff to absorb. The Bitcoin Red Team is shipping the bugs to the right people; the people on the other end are the rate-limiting step.
The next test is whether Western vendors tighten, loosen, or re-scope their dual-use rules for security research, and whether maintainers publish enough of the post-disclosure fixes to show the audit pipeline is producing patches, not just findings. The August volume is the first signal. The signal after that is whether the projects get quieter, or louder.