In 27.5 hours during the first week of August 2026, sixteen security researchers and three automated AI agents scanned 390 open-source Bitcoin repositories and filed 4,962 vulnerability findings — 85 rated critical, 635 rated high-severity. The effort, organized by the volunteer Bitcoin Red Team ...
"Discovery scales faster than responsibility. That single asymmetry is driving a crisis." — Mitchell Amador, CEO, Immunefi
In 27.5 hours during the first week of August 2026, sixteen security researchers and three automated AI agents scanned 390 open-source Bitcoin repositories and filed 4,962 vulnerability findings — 85 rated critical, 635 rated high-severity. The effort, organized by the volunteer Bitcoin Red Team and funded by a $40,000 OpenSats grant, produced in a single weekend what traditional security audit firms typically deliver in months.
The audit was not a theoretical exercise. It followed a chain of real-world incidents that have cost Bitcoin holders more than $130 million in confirmed losses since July 30, 2026. A five-year-old firmware flaw in Coinkite's Coldcard hardware wallet — apparently discovered by an attacker using frontier AI to review open-source code — enabled the systematic draining of at least 2,055 BTC from over 5,200 addresses. On August 3, non-custodial swap provider Boltz suspended all Bitcoin mainchain, Lightning, and Liquid swap services indefinitely, stating that "attackers now iterate faster than a team our size can find and patch." On August 28, Core Lightning shipped an emergency binary release (v26.06.7) fixing multiple vulnerabilities surfaced through AI-generated CVE reports, withholding source code for 14 days to give node operators time to upgrade.
Taken together, these events mark a structural shift in the economics of Bitcoin security: AI has driven the cost of vulnerability discovery toward zero while the cost of remediation — human coordination, code review, user migration — remains high and slow. The asymmetry is widening.
On the morning of July 30, 2026, a single attacker emptied 1,196 Bitcoin addresses in approximately 41 minutes. Roughly 1,082.65 BTC — worth approximately $70.2 million at the time — moved from wallets that had never touched the internet into a single consolidation address.
The attack exploited a firmware flaw dating to March 1, 2021, when a code migration in Coinkite's Coldcard hardware wallet silently rerouted seed generation from the device's hardware random-number generator to MicroPython's software fallback — an algorithm called Yasmarang, designed for devices with no randomness chip. The result was a collapse in effective key strength from a designed 128 bits to as little as 40 bits on affected devices — low enough to brute-force without physical access.
According to Galaxy Research, losses escalated rapidly across multiple waves:
| Wave | Date | BTC Stolen | Addresses Hit | |------|------|-----------|---------------| | First | July 30 | ~1,083 BTC | ~1,196 | | Second | Aug 1-2 | ~284 BTC | ~1,500+ | | Third | Aug 3-4 | ~229 BTC | ~2,500+ | | Suspected Fourth | Aug 4-7 | ~459 BTC (est.) | ~5,200+ | | Total | | ~2,055 BTC | ~5,200+ |
By August 4, TechCrunch reported confirmed losses exceeding $130 million. At least 15 separate attackers piled in after the initial exploit, according to Coinkite. The affected wallets belonged largely to long-term holders who generated recovery seeds using vulnerable firmware released from March 2021 onward.
Coinkite stated the company assumes "someone used frontier AI to review older firmware and uncover the issue." The wallet's source code is publicly available. On August 21, Coinkite shipped a new firmware update and confirmed the review behind it was AI-assisted, using Kimi and other frontier models to examine the entire codebase — not just the original randomness flaw. That review found additional issues in transaction approval, USB data handling, and firmware update validation.
According to PYMNTS, the incident demonstrated that AI is "collapsing the window between vulnerability creation, discovery, and exploitation."
On August 3, 2026, Boltz — a non-custodial Bitcoin swap service operating across Bitcoin mainchain, Lightning Network, and Liquid sidechain — suspended all swap operations indefinitely. The company cited months of escalating AI-assisted attacks that overwhelmed its development team.
Boltz's statement was direct: "Attackers now iterate faster than a team our size can find and patch."
The service had experienced a steady rise in automated, AI-assisted probing of its infrastructure. The pace of attacks accelerated sharply in the days before shutdown, forcing Boltz to conclude it could no longer operate safely. No user funds were lost — Boltz's HTLC-based non-custodial architecture protected customer assets — but the company absorbed operational losses on its own balance sheet.
The shutdown had immediate downstream effects. ZEUS, Aqua, and Bull Bitcoin — wallets and services that relied on Boltz infrastructure for swap functionality — lost access to atomic swap capability. By August 17, Blockstream had launched a beta swap service to partially fill the gap.
As of August 21, Boltz swaps remain offline. The original founders have stepped away, transferring the project to a group described as "Bitcoin veterans." No relaunch date has been announced.
The Boltz case illustrates a specific economic failure mode: for small open-source teams, the marginal cost of AI-powered attack discovery has fallen below the marginal cost of human-powered defense. The service was not poorly built — its non-custodial design prevented user fund losses — but the velocity of AI-discovered vulnerabilities made continued operation untenable.
On August 28, 2026, Blockstream released Core Lightning v26.06.7, an emergency point release addressing multiple vulnerabilities discovered through AI-generated CVE reports submitted over the preceding three weeks. In an unusual step, the team released signed binaries immediately but withheld the source code for 14 days, with publication scheduled for September 11, 2026.
The project has not disclosed what the vulnerabilities allow an attacker to do, how many are being fixed, or whether anyone has exploited them. No confirmed fund losses have been reported.
Core Lightning instructed node operators to upgrade immediately or, if unable, to restart their nodes with the --offline flag — which stops the node from communicating with other Lightning nodes while maintaining chain monitoring to detect channel counterparty cheating attempts. The project explicitly warned operators not to shut down nodes entirely: "It keeps running, which means it keeps watching the chain and can still act if a counterparty force-closes a channel."
According to the Blockstream blog post, more than 15 security reporters contributed to identifying the issues, including erickcestari, project-loupe, instagibbs, benthecarman, 0xaudron, callebtc, and others. A remediation team of nine developers, including rustyrussell and cdecker, prepared the fixes.
The 14-day source code embargo represents a calculated trade-off: giving operators an upgrade window while reducing the risk that technical details enable exploitation of unpatched nodes. The approach is borrowed from conventional software security practice but is relatively uncommon in Bitcoin's open-source culture, where code transparency is a foundational value.
Blockstream stated the project plans to "actively embrace AI-assisted review" for future releases.
The Bitcoin Red Team sprint, held August 4-5, 2026, was the largest coordinated AI-assisted security audit of Bitcoin open-source infrastructure to date. Sixteen globally distributed researchers and three automated agents scanned 390 repositories in 27.5 hours, averaging 180 findings per hour. Ninety-one percent of findings were surfaced through automated AI scans.
The severity breakdown:
| Severity | Count | % of Total | |----------|-------|-----------| | Critical | 85 | 1.7% | | High | 635 | 12.8% | | Medium/Low/Info | 4,242 | 85.5% | | Total | 4,962 | 100% |
The audit was funded by $40,000 from OpenSats. All issues were reported privately before public disclosure. The team plans to open-source its custom AI security harness, enabling Bitcoin companies to test closed-source software as well.
According to CryptoBriefing, the initiative was triggered directly by the Coldcard exploit. The Bitcoin Red Team, which includes pseudonymous developer Calle (a Bitcoin Red Team member), recognized that if AI could find a five-year-old firmware flaw in one project, it could find comparable flaws across hundreds of others.
The data suggests a troubling conclusion for the broader Bitcoin ecosystem: 720 high-severity or critical findings across 390 projects implies an average of nearly two serious vulnerabilities per repository examined. Many of these projects underpin wallets, exchanges, Lightning implementations, and other infrastructure handling real user funds.
Mitchell Amador, CEO of Immunefi — the largest crypto bug bounty platform — described the current environment as a "vulnerability apocalypse" at the WAIB Summit in Monaco. His data supports the characterization:
The pattern is consistent: AI has compressed the timeline and cost of vulnerability discovery by orders of magnitude, but human-driven remediation has not scaled proportionally. As Amador wrote, "Every item in that backlog is a vulnerability that someone might find and exploit."
The cURL project — one of the most widely used open-source libraries in the world — closed its bug bounty program entirely after AI-generated reports consumed maintainer time. This is not a Bitcoin-specific problem. It is a structural challenge for all open-source software that relies on small maintainer teams.
The events of July-August 2026 expose a structural economic problem in Bitcoin's security model.
The cost asymmetry is stark. The Bitcoin Red Team spent $40,000 on AI compute and 27.5 hours of human time to surface 4,962 findings. By comparison, a traditional security audit of a single major protocol costs $200,000-$500,000 and takes 4-8 weeks to produce a fraction of that output. The implication: any motivated attacker with access to frontier AI models and modest compute budget can audit Bitcoin open-source code at industrial scale.
Small teams cannot compete. Boltz's shutdown is the first confirmed case of an operational Bitcoin service becoming economically unviable due to AI-accelerated attack discovery. It will likely not be the last. The Lightning Network ecosystem in particular relies on small, often volunteer-maintained projects. As of May 2026, the network held approximately 4,898 BTC in public channel capacity across 41,080 channels and 17,438 nodes — but the capital is concentrating into fewer, larger, better-run nodes. Monthly payment volume exceeds $1.1 billion. The infrastructure supporting that volume is maintained by a small number of teams.
Static audits are obsolete. The Coldcard flaw existed for five years in publicly available source code. It passed multiple human reviews. An AI model found it. As PYMNTS reported, "Code that passed an audit three years ago may need to be continually reexamined as models become better at understanding complex systems." The shift from periodic auditing to continuous AI-assisted review is not optional — it is a competitive requirement for any project holding user funds.
Bug bounties are under strain. Immunefi processed 24,418 submissions in 2025 — a 135% increase. At current growth rates, platforms designed for human-scale vulnerability reporting may be overwhelmed by AI-generated submissions within 12-18 months. The signal-to-noise problem compounds: more reports does not mean more actionable findings, and triage itself becomes a bottleneck.
The defender's advantage is eroding. In traditional security, defenders have structural advantages — they know their own systems, they can patch before disclosure, and they benefit from coordination. AI compresses the exploit window so aggressively that these advantages are diminishing. When a firmware flaw can be discovered, weaponized, and deployed across thousands of wallets in hours rather than months, the entire responsible disclosure model comes under pressure.
The events of July-August 2026 do not represent a failure of Bitcoin's cryptographic foundations. The underlying mathematics remain sound. What has failed — or more precisely, what has been outpaced — is the human-scale security infrastructure surrounding it: the firmware reviews, the code audits, the bug bounty programs, the small maintainer teams.
AI has not broken Bitcoin. It has broken the assumption that open-source code review at human speed is sufficient to secure billions of dollars in user funds. The Coldcard exploit demonstrated that a single AI-discovered flaw in a hardware wallet can drain $130 million. The Boltz shutdown demonstrated that a small team can be rendered operationally unviable by AI-accelerated attacks. The Core Lightning embargo demonstrated that even well-resourced projects must resort to withholding source code — an act antithetical to open-source culture — to buy time for human-speed remediation.
The Bitcoin Red Team's 4,962 findings in 27.5 hours suggest the problem extends far beyond isolated incidents. With 720 high-severity or critical findings across 390 projects, the backlog of unresolved vulnerabilities in Bitcoin infrastructure is substantial and growing faster than the community can address it.
The economic logic is clear: projects that do not adopt continuous AI-assisted security review face escalating risk. The cost of inaction is no longer measured in theoretical exposure — it is measured in the $130 million already lost from Coldcard wallets and the indefinite suspension of services like Boltz. For the Lightning Network's $1.1 billion monthly payment volume and the broader Bitcoin ecosystem, the question is not whether AI-assisted security becomes standard practice, but whether adoption occurs fast enough to prevent further losses.