Bitcoin Core Developers Debate AI Security After Coldcard Hack and AI Bug-Report Flood
From Coldcard to Bitcoin Core: AI Raises the Stakes for Crypto Security
The July 2026 Coldcard hardware wallet incident exposed a five-year-old firmware flaw that reduced seed entropy far below the intended 128 bits, allowing attackers to reconstruct private keys without physical access to devices. Losses climbed past 1,778 BTC, valued at roughly $112 million to $130 million depending on the measurement window, across more than 8,000 addresses in multiple waves that began on July 30. Coinkite, the manufacturer, stated that the open-source nature of the firmware made AI-assisted review a plausible discovery method for the bug, which had sat undetected despite prior human and automated checks. In the weeks that followed, a volunteer effort known as the Bitcoin Red Team used frontier and open-weight models to scan 390 Bitcoin-related repositories, producing 4,962 findings that included 85 critical and 635 high-severity issues in approximately 30 hours.
Parallel pressure appeared in other infrastructure: Core Lightning developers issued emergency guidance after a flood of AI-generated reports confirmed real vulnerabilities. Against this backdrop, Bitcoin Core contributors convened their weekly meeting in early September 2026 to examine whether the project had fallen behind in both defending against AI-enabled threats and employing AI for code review, testing, and issue triage. The Coldcard exploit and the subsequent wave of AI-assisted vulnerability discovery have forced Bitcoin Core developers and the wider ecosystem to confront a structural shift in security economics, where the speed and scale of machine analysis outpace traditional human review, yet consensus on formal AI policies remains elusive because the project operates as a decentralized collective rather than a centralized organization.
Coldcard Firmware Entropy Collapse Traced to 2021 Build Configuration Error
The underlying issue that led to the Coldcard losses can be traced back to a firmware release in March 2021, specifically version 4.0.1 and its subsequent builds. This release inadvertently caused the seed generation process to revert to a deterministic software random-number generator, rather than utilizing the device’s intended hardware entropy source. As a result, on the affected Mk2 and Mk3 devices, the effective entropy significantly decreased to approximately 40 bits, while newer models managed to retain around 72 bits of entropy. Beginning early on July 30, 2026, attackers systematically exploited this vulnerability by regenerating the weakened seeds and subsequently sweeping funds from single-signature addresses. Galaxy Research conducted an investigation and documented high-confidence losses amounting to 1,778.84 BTC across more than 8,600 addresses, following direct contact with 190 victims.
Notably, the majority of the stolen coins remained untouched in addresses controlled by the attackers for several weeks. Coinkite confirmed that the firmware updates implemented only safeguard newly generated seeds, indicating that existing compromised seeds necessitate a complete migration to new entropy sources to ensure security. The company publicly speculated that the review of the long-public source code by AI had likely expedited the discovery of the latent logic error that human auditors had overlooked for five years. This sequence of events illustrated that while open-source transparency has long been regarded as a security advantage under the principle of “don’t trust, verify,” it also inadvertently expands the attack surface, especially when models can analyze entire repositories in mere minutes.
Bitcoin Red Team Completes Rapid AI-Assisted Audit of 390 Repositories
In the immediate aftermath of the Coldcard waves, a dedicated volunteer group comprising approximately 16 to 25 skilled developers and researchers, coordinated by prominent figures including Calle and Rob Hamilton of AnchorWatch, initiated a comprehensive and coordinated AI-driven review of the Bitcoin ecosystem code. Over the course of 27.5 to 30 hours, this extensive effort resulted in the filing of an impressive total of 4,962 findings across 390 distinct projects. Among these findings, 85 were initially rated as critical, while 635 were classified as high severity, leading to an average of more than two serious issues identified per person-hour of work. The models employed during this review included Kimi K3, GLM 5.2, and several others, particularly after access limitations emerged with certain frontier providers.
Notably, privacy and coinjoin-related projects exhibited the highest concentration of serious findings, highlighting the vulnerabilities present in these areas. The campaign operated under a dual process that involved automated scanning followed by meticulous human reproduction of the findings and private disclosure to maintainers prior to any public release of the information. It is worth mentioning that eight findings were later retired as false positives, which illustrates the ongoing and residual need for thorough verification in such audits. The output generated by the Red Team provided concrete data points that later played a significant role in informing Bitcoin Core discussions regarding the practical value and noise level associated with AI-generated reports, ultimately contributing to a deeper understanding of the implications of such technology in the ecosystem.
Core Lightning Faces Emergency Response to AI-Generated Vulnerability Reports
Core Lightning developers recently reported a significant and unexpected influx of AI-generated security submissions, which necessitated an extensive triage process lasting ten days. This effort was undertaken by a small dedicated team alongside several outside contributors who volunteered their time and expertise. During this period, a number of the submitted reports were found to be valid, which ultimately led to the urgent need for an emergency security release to address the identified vulnerabilities. As a precautionary measure, operators were promptly instructed to install the verified updates to ensure the security of their systems. Alternatively, they were advised to restart their nodes using the offline flag, allowing them to continue monitoring the blockchain while simultaneously blocking any new payments and peer connections to mitigate potential risks.
In light of these developments, the project announced that it would no longer provide support for certain earlier versions of the software, signaling a shift in focus towards more secure and updated iterations. A subsequent major release was scheduled for late September 2026, indicating a proactive approach to enhancing the overall security framework. This served to illustrate a critical point: the very same tools and technologies that expedite the discovery of vulnerabilities also contribute to an increased triage overhead for open-source maintainers who are often under-resourced. This challenge is not unique to Core Lightning; similar patterns have been observed in other layer-two and sidechain codebases. These observations reinforce the notion that as complexity increases outside the foundational Bitcoin protocol, the associated risks become more concentrated, particularly when AI is capable of scanning and analyzing data at machine speed.
Bitcoin Core Weekly Meeting Centers on AI Adoption Gap
At the weekly Bitcoin Core developer meeting held the first Thursday of September 2026, Spiral lead Steve Lee relayed donor and community concerns that the project had neither sufficiently hardened against AI-enabled network threats nor systematically applied AI to code review, testing, and debugging. Protocol Watch founder Christine D. Kim summarized the discussion publicly, noting that recent losses exceeding 1,800 BTC from more than 5,200 addresses and the Core Lightning report surge had raised the stakes.
Access Restrictions on Frontier Models Shape Defensive Capabilities
Researchers engaged in the Red Team and associated initiatives reported that certain frontier model providers based in the United States imposed restrictions on cybersecurity-oriented queries. This situation forced cybersecurity professionals to rely on open-weight alternatives, such as Kimi K3, which do not have the same limitations. Galaxy Research conducted an assessment and concluded with high confidence that at least some malicious actors had utilized unrestricted models that possess similar capabilities to those that were restricted. This created a significant asymmetry in the cybersecurity space, where attackers were able to query and leverage powerful systems, while some defenders found themselves hindered by policy blocks. This disparity became a recurring theme in the analysis conducted after various incidents.
In response to these challenges, industry participants, including several organizations focused on Bitcoin and cryptocurrency, later advocated for the establishment of trusted-access programs. These programs would grant vetted maintainers of open-source financial infrastructure broader capabilities with respect to model usage, but only under controlled conditions to ensure safety and security. The ongoing debate highlighted a critical issue: the safety policies implemented by model providers, which are designed to reduce general risk, can inadvertently produce uneven effects across different categories of legitimate security research. This situation calls for a careful reconsideration of how such policies are structured and enforced to ensure that they do not stifle important research and defensive capabilities in the cybersecurity domain.
Rob Hamilton Lands First Bitcoin Core Commit from AI-Assisted Discovery
AnchorWatch CEO Rob Hamilton got his very first merged commit into the Bitcoin Core codebase on August 20, 2026, as an active participant in the Red Team initiative. Specifically, this major update fixed a major wallet crash bug uncovered by advanced AI models while scanning the massive codebase for problems. Hamilton had started to experiment with AI-assisted analysis of Bitcoin Core as early as May, but the Coldcard incident noticeably catalyzed this into a well-coordinated campaign to improve the software. The successful merge of this commit demonstrated that findings derived from AI can indeed translate into accepted and meaningful protocol-level improvements, particularly when these findings are accompanied by careful human validation and thorough review processes.
At the same time, Hamilton and his collaborators observed that the primary bottleneck in the development process had shifted from the initial discovery of issues to the considerably slower and a more complex process of coordinating with maintainers across hundreds of different repositories. They found that identifying thousands of potential issues in a single day proved to be far easier than securing timely fixes and achieving independent reproduction of the identified problems. This shift in focus showed the challenges faced by developers in the open-source community as they worked to implement necessary changes and improvements.
Entropy and Random-Number Generation Reexamined Across Hardware Wallets
The post-mortem analysis of the Coldcard incident prompted a significant response from manufacturers and independent reviewers alike, leading them to undertake a thorough re-audit of the random-number generation paths utilized in various other devices. In response to these developments, Blockstream took the initiative to publish a comprehensive analysis that confirmed its Jade wallet remained unaffected by the issues raised and provided detailed insights into the prior improvements made to its own entropy handling processes. Coinkite, the manufacturer of Coldcard, released a new firmware update that mandates that users supply physical sources of randomness, such as key presses, die rolls, or even coin flips, when generating new seeds. This update effectively replaced the previous software fallback algorithm that had been in place.
This incident served to underscore a critical point: the mere presence of a hardware random-number generator (RNG) component does not inherently guarantee that it will be invoked correctly during operations. It is essential that the entire call path, starting from seed generation and extending through to build configuration, is meticulously verified from end to end. Following this incident, subsequent AI scans were conducted across multiple hardware-wallet repositories and yielded additional reports, which were then triaged by maintainers. This process illustrated how a single high-profile failure can lead to a cascade of scrutiny throughout the entire ecosystem, prompting a reevaluation of security practices and protocols across the board.
Open-Source Transparency Meets Machine-Speed Analysis
Jameson Lopp and other long-time self-custody advocates observed that the Coldcard case exposed practical limits of the long-standing “don’t trust, verify” maxim. Human reviewers had examined the relevant firmware for years and confirmed the existence of the hardware RNG, yet the silent fallback path remained undetected until models were pointed at the repository. Coinkite’s own earlier AI review of the same code had also missed the issue.
The perspective does not invalidate open-source practices; rather, it demonstrates that verification workloads must now incorporate continuous machine analysis calibrated to the same models available to potential adversaries. Projects that maintain public repositories face a new baseline expectation: assume that every commit is already under automated scrutiny by both defenders and attackers.
Layer Two and Sidechain Codebases Attract Concentrated AI Attention
In addition to hardware wallets, artificial intelligence scans and the subsequent exploits have increasingly targeted more intricate components of the Bitcoin stack. Notably, Core Lightning, the Liquid Network, and a variety of payment and privacy tools have all encountered heightened volumes of reports or direct incidents related to vulnerabilities. One detailed analysis pointed out that while the foundational Bitcoin protocol is intentionally designed to remain straightforward and uncomplicated, the surrounding infrastructure that is necessary for enhancing usability and scaling capabilities introduces significantly larger attack surfaces. These surfaces are ones that advanced models can navigate with remarkable efficiency.
White-hat activities conducted on the Liquid Network led to the temporary withdrawal of a substantial amount of value, which was subsequently returned after appropriate patching measures were implemented. Collectively, these events indicate a noteworthy shift in the economic calculus surrounding vulnerability research: the marginal cost associated with scanning additional repositories continues to decrease, while the potential financial payoff from discovering a single high-value bug remains considerably substantial.
Coordination Challenges Persist Despite Faster Discovery Rates
Despite the significant advancements made by artificial intelligence in reducing the time required to identify potential issues, from what used to take weeks or even months down to mere hours, the human processes involved in triage, reproduction, private disclosure, and the integration of patches have continued to follow their previous rhythms and timelines. This has led to challenges, particularly for maintainers of smaller projects, who have reported experiencing difficulties in managing the sudden and substantial influx of reports that have emerged as a result of these technological improvements.
The Red Team’s established practice of private disclosure for critical findings has effectively mitigated immediate risks to the public; however, it still necessitated a sequential engagement process with dozens of independent teams, which can be quite resource-intensive and time-consuming. OpenAI’s temporary restrictions placed on certain accounts have further complicated the continuity of tooling and processes that are essential for effective project management. The cumulative effect of these factors has resulted in a widening gap between the speed at which vulnerabilities are discovered and the speed at which they can be remediated. This situation represents an organizational challenge rather than a purely technical one, and it is one that the decentralized structure of Bitcoin Core is still in the process of adapting to effectively address.
Industry Calls for Trusted Access to Advanced Models Intensify
In light of the identified disparities, a coalition of various organizations aligned with Bitcoin has come together to endorse and formally request the establishment of ongoing trusted-access programs at prominent artificial intelligence laboratories. These proposals put forth the argument that individuals who are recognized as qualified defenders of open-source financial infrastructure ought to be granted access to capability levels that are on par with those that are typically available for general research purposes, all while ensuring that safety guardrails are maintained and upheld.
In the absence of such access channels, these defenders face the significant risk of falling behind systematically when compared to attackers, who are subject to far fewer restrictions regarding their usage. As of early September 2026, this discussion continues to be a vibrant and active topic, intersecting with broader policy dialogues that address the dual-use nature of advanced language models and their implications for various sectors.
Individual Contributor Practices Continue to Drive AI Integration
Due to the absence of a corporate hierarchy within Bitcoin Core, the process of adopting artificial intelligence unfolds through the decisions made by individual developers rather than through directives imposed from above. A number of contributors have already begun to depend significantly on various models for purposes such as test generation, identifying fuzzing targets, and conducting preliminary reviews; however, there are others who remain doubtful regarding the quality of the output produced by these models and thus prefer to minimize their exposure to what they perceive as generated noise.
During the meeting held in September, participants left open the possibility that a social consensus might eventually emerge to support the establishment of project-level guidelines. Nevertheless, those involved treated this potential outcome as something that is contingent upon various factors rather than something that is imminent or guaranteed. In the meantime, the wide-ranging diversity of individual practices serves a dual purpose: it acts as a strength by fostering an environment conducive to experimentation while simultaneously being a source of uneven coverage throughout the codebase, leading to inconsistencies in implementation and application.
Practical Impact for Self-Custody and Multi-Signature Setups
Users who generated seeds on affected Coldcard firmware were advised to migrate immediately to new entropy sources and, where feasible, to multi-signature configurations spanning devices from independent manufacturers. No confirmed thefts from properly constructed multi-signature wallets were reported in the high-confidence tallies.
The incident reinforced long-standing guidance that single points of failure, whether a single device model or a single generation process, carry outsized risk once discovery tools improve. Hardware wallet makers responded with firmware changes that shift entropy collection toward user-supplied physical randomness, reducing reliance on internal generators whose correctness is harder to audit at a glance.
🔥 Beyond the Headlines: What KuCoin 5.0 Means for You
Market news moves fast — but where you act on it matters just as much. This October, KuCoin launches KuCoin 5.0, transforming KuCoin into a rebuilt platform. Here's what actually changes for you:
-
One account for everything. Older platforms split your money across separate "spot," "margin," and "futures" accounts and expected you to understand why. KuCoin 5.0's unified account removes that entirely — deposit once, and everything is simply there.
-
Stocks, indices, and commodities. KuCoin 5.0 expands beyond crypto into global markets. When crypto chops sideways and equities rally (or the reverse), you rotate in minutes instead of opening a brokerage account and waiting days for fiat rails.
-
Real-world assets (RWA). Tokenized exposure to traditional assets like commodities, right inside your crypto account. One of the fastest-growing segments in global finance is no longer reserved for institutions — you access it from the same balance you trade with.
-
Earn while you learn. Not ready to trade? KCUSD lets your stablecoins earn daily, auto-compounding interest. The lowest-stress way to put your idle deposit to work for 4% yield.
-
An AI assistant in plain language. Ask questions, get market context, understand what you're looking at — built into the platform, no jargon required.
-
An app that doesn't overwhelm. Faster, cleaner, and consistent — intuitive from the first tap, not after a tutorial.
-
Safety you can check, not just trust. A MiCAR-licensed EU entity, Proof of Reserves you can verify yourself, and internationally certified security (SOC 2 Type II, ISO 27001:2022).
Create your account in minutes — and start on the platform built for where crypto is going, not where it's been.
FAQs
How did the Coldcard vulnerability remain undetected for five years despite open-source availability?
Human reviewers confirmed the presence of the hardware random-number generator but did not exhaustively trace every execution path from seed generation through the specific build-configuration flags that silently activated the software fallback. AI models, when directed at the full repository, surfaced the logic error more quickly by correlating distant code sections that individual auditors had examined in isolation. Coinkite’s own prior AI review also failed to flag the issue, underscoring that model capability and prompt design remain variable.
What proportion of the Bitcoin Red Team findings were ultimately actionable?
Of the 4,962 initial findings, 85 critical and 635 high-severity reports constituted the core set of serious issues. Approximately 21 percent of reported issues had been independently reproduced at intermediate checkpoints. Eight findings were retired as false positives. The remainder required ongoing maintainer triage, illustrating that volume alone does not equal validated risk.
Did attackers definitively use AI to discover the Coldcard flaw?
Coinkite stated that the open-source firmware made AI review a plausible discovery vector and that its own earlier model scans had missed the bug. Galaxy Research assessed with high confidence that at least some attackers employed unrestricted models during exploitation. Direct forensic proof of the initial discovery method remains incomplete, yet the subsequent wave of AI-driven audits across the ecosystem demonstrated that the capability exists and is already in active use by both sides.
Why did Bitcoin Core developers decline to adopt a formal AI policy?
Participants at the September meeting emphasized that the project is not a company and therefore cannot impose uniform technical standards on independent contributors. Existing usage already varies widely by individual preference. Formal policy was left contingent on the emergence of broader social consensus within the developer community.
How are Core Lightning operators expected to respond to the recent security releases?
Operators were instructed to install and verify the emergency update promptly. Those unable to upgrade immediately were advised to restart with the offline flag, allowing the node to continue monitoring the Bitcoin chain while refusing new payments and peer connections. Support for certain older versions was discontinued.
Disclaimer: This content is for informational purposes only and does not constitute investment advice. Cryptocurrency investments carry risk. Please do your own research (DYOR).