6,700 findings. 1,029 flagged high or critical. 55 hours. 425 repositories.
Those numbers are already circulating through Bitcoin Twitter. The trigger: Bitcoin Red Team, a volunteer security sprint that swept the ecosystem's public codebase using a fleet of frontier LLMs. On paper, the output is staggering. Roughly 21 humans, assisted by Kimi K3, GPT Sol, Fable/Opus, GLM 5.2, and OpenAI's Cyber Harness, produced more candidate vulnerabilities in under three days than most audit firms publish in a year.
Here's the problem: nobody knows what those numbers mean. No denominator. No false-positive rate. No reproduction steps. No model versions. Just raw aggregates.
I've watched this industry long enough to distrust impressive counts. In late 2017, I audited 15 ERC-20 whitepapers for an angel syndicate and flagged a reentrancy vulnerability in EtherStatus before its mainnet launch. The team pulled $200,000. Two weeks later, the project rugged. Raw findings are candidates, not conclusions. That distinction has cost people everything.
Context
Bitcoin Red Team is not a company. No token. No TGE. No treasury. No incentive design. It's a flash event organized by Rob Hamilton, a security researcher whose work on the Coldcard hardware wallet incident appears to have served as the catalyst. The playbook: pull public repositories across the Bitcoin ecosystem, run them through a multi-LLM pipeline, generate candidate findings, then hand those candidates to domain experts for reproduction and triage.
The first snapshot: 150 repositories, 6,700 findings, 1,029 high or critical, at roughly $20,000 in compute cost. The second snapshot pushed coverage to 425 repositories. Per-repo scan cost: roughly $130-$150. That's cheap. Almost trivially cheap for the coverage.
Compare that to traditional audit houses. Trail of Bits and CertiK charge per project, per scope, with delivery cycles measured in weeks. They produce formal reports, reproducible findings, and named reviewers. What they don't produce is ecosystem-scale coverage. The fastest traditional audit covers one codebase. This sprint covered 425 in 55 hours. The trade is obvious: depth for breadth. Whether that trade is worth it depends entirely on what the triage stage produces.
The methodology is a layered pipeline. The models do broad-pattern hunting. The experts shape prompts, interpret outputs, attempt reproduction, and decide what gets disclosed. On paper, that sounds reasonable. In practice, it's a funnel: AI generates candidates at scale, humans validate at bottleneck speed.
Here's the uncomfortable structural fact: this is not a new paradigm in vulnerability research. It's a process innovation, scale and speed grafted onto a traditional manual-audit workflow. Process innovations have failure modes that technology innovations don't. They inherit the biases of their human operators. And those biases — what experts choose to chase, what they choose to publish — are invisible to anyone watching from the outside.
Core
Rob Hamilton was explicit about the bottleneck: it's not GPU time or API credits. It's operations, disclosure handoff, and triage. In plain terms: AI scanning now runs faster than qualified humans can review. That's the real headline.
This mirrors my own experience. In 2026, I integrated AI-driven sentiment analysis into our quantitative stack, 10,000 news articles daily feeding model adjustments. The system identified a 5% alpha edge during low-volume windows. Then it misread a geopolitical headline, and I manually halted trading to prevent a $500,000 loss. Machines generate at scale. Humans validate at bottleneck speed. Always.
The Bitcoin Red Team methodology is explicitly human-in-the-loop. Hamilton's own words: a domain expert can change a finding's severity with a sentence or two of context. A snippet of protocol knowledge pushes a medium to critical. That's not autonomous auditing. That's AI-powered triage feeding human judgment.
The participant data confirms this. At 27.5 hours, 16 participants. At 55 hours, 24 participants, three of them bots. Effectively 21 humans. Growth of roughly five people over 28 hours. This event is constrained by the availability of qualified security researchers, not compute. If you scale this model, you need more experts, not more GPUs.
The economics are the tell. Compute costs run $130-$150 per repository. But the bottleneck is human. Every hour of expert triage costs more than the entire scanning bill for dozens of repositories. The true per-finding cost is multiples of the advertised number. Traditional audit firms charge $50,000-$500,000 per engagement because the human labor is the product. This sprint concentrates the human labor differently, but it doesn't eliminate it. AI didn't make security review cheap. It made the cheap part faster and left the expensive part untouched.
Now the numbers. The early snapshot averages about 44 findings per repository. But vulnerability distributions are never uniform. The realistic structure: a handful of repositories contribute a disproportionate share of findings, while most yield nothing actionable. Without per-repo breakdowns, the aggregate counts are close to meaningless for risk assessment.
The single most important statistic in this entire story: only 19.5% of scanned Bitcoin ecosystem projects have a SECURITY.md file. That's not a vulnerability count. That's an infrastructure gap. The disclosure infrastructure is largely nonexistent across the ecosystem. An external scanner walking in uninvited is filling a hole that should have been institutionalized years ago.
The event has produced dozens of disclosures so far, under 10% of scanned projects at the 150-repository mark. Extrapolate to 425 and you get perhaps 20-40 total disclosures if the trend holds. That's the realistic output envelope. The other 6,600-plus findings are untriaged candidates sitting in a queue.
Contrarian
The market will misprice this event. The 6,700 figure is FUD ammunition. Expect the narrative within days: Bitcoin L2s are leaking everywhere. Expect competing security firms to cite this data to justify larger mandates. That's how this industry eats.
But the contrarian truth: this event is net-positive for Bitcoin security. Even if only 20 real critical vulnerabilities get fixed because an underfunded volunteer sweep covered repositories no audit firm would touch, that's a public good. Bitcoin's security culture has historically centered on consensus-layer cryptography, not application-layer audit hygiene. This sprint creates the first broad-funnel scan of the ecosystem. Flawed. Non-reproducible. But existentially valuable as a starting point.
The real danger isn't the scanning. It's the disclosure process. The event materials mention immediately disclosing severe findings once a proof-of-concept confirms exploitability. Immediate is the wrong word in responsible disclosure. Maintainers need a window to patch, industry standard is 90 days. If immediate means public disclosure, this becomes a 0-day broadcast mechanism. The legal exposure alone is nontrivial: unauthorized access statutes in most jurisdictions don't care about your intent once you move from static analysis to dynamic exploit validation.
One more signal worth watching: the Coldcard research that catalyzed this effort hasn't been attributed to the sprint itself. The organizers appear to have done hardware-level work before this event went live. If this sprint is an extension of one researcher's existing work rather than a general-purpose engine, the reproducibility of the model is even lower than the raw numbers suggest.
The uncomfortable question: what happens to the 98% of findings that never get triaged? No one knows. That's the honest answer. This event generates candidates, not conclusions. The raw count is a marketing number. The disclosures are the only verifiable output, and they number in the dozens.
Takeaway
Data speaks, but only if you know how to listen. The market will treat 6,700 as a threat vector. Smart money will wait for the disclosure log.
The next 90 days determine whether this is a template or a footnote. Confirmations. Patches. Public tracking. If the validation rate holds, this becomes the prototype for ecosystem-wide AI-assisted security sprints. If it collapses, it's another dataset buried in a Telegram channel.
Operator checklist: demand the denominator. Ask for per-repo severity distributions. Require reproduction steps before acting. And for the love of diligence, confirm your project has a SECURITY.md.
Due diligence is the only hedge you control. Right now, 80.5% of the Bitcoin ecosystem is running naked.
Ledgers do not forgive, they only record. This event just recorded 6,700 entries. The question is how many survive contact with reality.
