
AI Identifies Nearly 5,000 Bitcoin Vulnerabilities in One Day: Chinese Large Models Show Strength, Open Source Does Not Equal Security
By: Xiao Bing
16 people, 27.5 hours, 4,962 security issues, 85 critical vulnerabilities, 635 high-severity vulnerabilities.
These numbers come from a volunteer organization called Bitcoin Red Team. Over the past two days, they launched a carpet-style security audit on 390 Bitcoin open-source projects using AI models, submitting an average of 166 findings per hour, with each person identifying one critical-level vulnerability per hour.
Cashu protocol founder Calle published the first battle report on X, summarizing the situation in the first four characters: "The situation is extremely bad."
But what is truly unsettling is not the numbers themselves.
$40,000 and a Fire Drill
This audit was not launched out of thin air.
A week ago, the Bitcoin ecosystem just experienced the largest security incident in hardware wallet history. Due to a random number generator flaw in a 2021 firmware update, Coldcard wallets allowed private keys to be calculated offline. Starting July 30, attackers looted over 5,200 addresses in four waves. Tracking data from Galaxy Research shows losses of approximately 1,816 BTC, calculated at over $116 million at the time.
This Bug sat quietly in the public code for five years; anyone could audit it, yet almost no one looked at it systematically.
The Coldcard incident was the trigger. Calle and AnchorWatch CEO Rob Hamilton quickly assembled the Red Team, secured funding from OpenSats, and burned through over $40,000 on AI compute. The models they used included Moonshot AI's Kimi K3, OpenAI's GPT Sol, Anthropic's Fable and Opus, and Zhipu AI's GLM5.2.
In the early stages of the project, access to OpenAI and Anthropic models was restricted, so the team relied mainly on Chinese open-source models. This detail itself is quite interesting: Those guarding the security of the world's largest crypto asset infrastructure turned out to be a group of volunteers wielding Chinese open-source AI fighting through the night.
The data puzzle after 27.5 hours looks like this: Of 4,962 issues, 14.5% were high or critical severity, averaging 1.85 critical issues per project. Privacy and CoinJoin-like tools had the highest proportion of critical findings at 24%; followed by exchanges and swap protocols at 21%; cryptography libraries and SDKs produced the highest volume of raw findings (1,101), but the critical proportion was only 10%. 91% of findings were submitted via automated scanning, and 21% were reproduced in local environments using proof-of-concept code.
Of the 17 contributors, 14 were humans and 3 were automated systems.
The Bottleneck Has Flipped
Rob Hamilton wrote a sentence on X that is more important than all the numbers: "The hardest part is not finding Bugs, but getting them into the right hands."
In the world of traditional security research, "discovery" is the most expensive phase. A top security researcher might spend weeks reverse-engineering a piece of code to find an exploitable vulnerability. AI has pressed this cost down to near zero. 16 people working with AI models can produce in one day what equals months of work for an audit firm.
The problem lies after discovery. As of the release of the battle report, only 19 out of 390 audited projects (less than 5%) had completed upstream disclosure. Calle apologized to maintainers in a tweet, saying they knew the surge of reports increased their pressure and the team was still learning how to filter noise. 8 findings had been withdrawn as false positives.
This is a structural change: AI has made "finding Bugs" a nearly free act, shifting the bottleneck of the entire security pipeline from the upstream discovery phase suddenly to the mid-to-downstream phases: verification, classification, routing to the correct maintainers, fixing, testing, and deploying patches. These phases still rely heavily on manual labor, remain slow, and remain chaotic.
And this is exactly where the attackers' structural advantage lies.
An Asymmetric Arms Race
Defenders need to complete a full chain: discover vulnerability → verify → responsible disclosure → wait for maintainer confirmation → fix → push patch → user update. Every step has friction and delay.
Attackers only need: discover vulnerability → exploit.
AI makes the starting point of the chain cheap for both sides. But the asymmetry of subsequent steps means this arms race naturally favors the offensive side. The Coldcard case is clear proof: a 2021 Bug sat publicly in the codebase for five years; defenders did not systematically scan, attackers used AI to find the predictable key space, and swept away 1,083 BTC in 41 minutes.
Larger signals have already emerged. In April this year, Anthropic showcased Claude Mythos Preview externally, an AI model refused public release due to excessive security risks. According to Anthropic, Mythos could lock onto a single vulnerability with less than $50 in compute costs; the team used it to autonomously discover a flaw in the OpenBSD operating system that had existed for 27 years. OpenBSD is one of the globally recognized most secure operating systems, widely used to run firewalls and critical infrastructure. Anthropic therefore launched the Project Glasswing initiative, partnering with about 40 organizations including AWS, Apple, Microsoft, and Google, to use this model to patch vulnerabilities before attackers.
In May, the Google Threat Intelligence Team announced another milestone: they intercepted a case where a criminal group used an AI model to discover a zero-day vulnerability and wrote exploit code. That vulnerability was a two-factor authentication bypass embedded in a widely used open-source management tool. Google identified the attack source through AI characteristics in the code (fictional CVSS scores, textbook-style Python comments) and collaborated with the vendor to complete the patch before the criminal group launched large-scale exploitation.
Google Threat Intelligence Chief Analyst John Hultquist's words left no room: "The view that the AI vulnerability arms race is coming is a misjudgment. The reality is, this race has already begun."
Open Source Never Equals Audited
Ledger CTO Charles Guillemet said a widely quoted sentence when evaluating the Coldcard incident: "Open source and being audited are two different things."
This sentence pierced a long-existing cognitive bubble in the crypto industry. The Bitcoin community has a nearly religious faith in open source, believing that public code means anyone can check it, thus it is naturally secure. Coldcard's code was indeed public for five years. The Red Team's audit covered 390 projects. Putting these two facts together, the picture is clear: Auditable never equals audited.
The Red Team's experience also exposed the limitations of AI security auditing. Calle said most of the team's work was still "manually guiding AI," letting everyone prompt the model in their preferred way, resulting in broader coverage than a unified method. This shows that current AI security auditing is closer to "human experts driving AI tools" rather than "AI autonomous patrolling." Models can quickly scan code patterns, but judging whether a finding is truly exploitable, the scope of impact, and who to notify still requires human experience and judgment.
Hamilton said the team plans to open-source the audit framework, allowing Bitcoin companies to run the same scans on their closed-source code. This is the right direction, but it also means the same tools will inevitably fall into the hands of attackers.
What this audit truly reveals is a new equation taking shape:
AI makes the cost of discovering vulnerabilities approach zero. But the cost of fixing vulnerabilities (organization coordination, manpower investment, user migration) remains expensive and slow.
$40,000 in AI compute, 27.5 hours, can turn an open-source ecosystem managing hundreds of billions of dollars in assets upside down. The Coldcard Bug that no one looked at for five years was found by attackers and swept away $116 million in 41 minutes.
Bitcoin's security narrative is being rewritten.
The past logic was: "The code is open source, so it is secure." The new logic is crueler: "The code is open source, so attackers are also using AI to scan it."
The defenders' only advantage is the first move, finding and patching vulnerabilities before attackers. The Red Team is fighting for this first move, but their battle report also shows that even with the first move, the subsequent patch distribution chain remains the weakest link in the ecosystem.
The dilemma facing Bitcoin may only be a preview of a larger change. The entire open-source software world is about to face the same questioning: When finding Bugs no longer requires experts, but fixing Bugs still does, what will fill this gap?













