AI models are moving deeper into crypto security, and Bitcoin is becoming one of the clearest battlegrounds.
The change is playing out on both sides. Attackers are using AI to cut the time and cost required to find and exploit vulnerabilities. Defenders are using it to scan code, identify weaknesses, reproduce bugs, and speed up fixes. For a Bitcoin ecosystem protecting more than $1 trillion in assets, that shift is no longer theoretical.
AI is speeding up offensive security work
Hardware wallet maker Coldcard was hit by a large-scale hack not long ago, sending shockwaves through the crypto community in China and overseas. According to the disclosure cited in the report, more than $100 million in Bitcoin was stolen. Beyond the direct losses, the incident also shook long-held confidence in the safety of Bitcoin self-custody.
Coldcard had long been treated as a high-security vault by many market participants. Because its code is open source, community members suspected the attacker may have used AI to review older firmware and identify a vulnerability that had remained hidden for five years across a large body of historical code. After the attack, the community used Claude Code and located the flaw in roughly eight minutes.
Another case emerged this month when non-custodial Bitcoin exchange service provider Boltz said it would suspend Bitcoin exchange services indefinitely. One reason it gave was that AI-assisted attacks were evolving faster than the team could fix vulnerabilities. Over the past few months, Boltz said it had been dealing with automated AI-driven probing and had already handled multiple exploitation incidents. More recently, the pace picked up, and the company suspected that several well-resourced attack groups were targeting the platform at the same time. With no way to guarantee safe operations during remediation, Boltz halted its swap service. The report said it was later taken over by an anonymous Bitcoin team.
As large models keep improving, attackers are weaponizing them. Tasks that once demanded substantial time for code review and exploit path design are being automated piece by piece. That means code can be analyzed faster, flaws can be identified at lower cost, and exploitation chains can be built with greater speed and concealment.
The report also points to North Korean hacking group Kimsuky, which has recently expanded its use of generative AI. In addition to producing phishing lures with AI, the group has built an independently operated local large language model environment and a retrieval-augmented generation, or RAG, system in an attempt to automate intelligence extraction and attack workflows more deeply. In practice, Kimsuky has begun using generative AI to create highly realistic documents tied to virtual assets and finance for targeted spear-phishing campaigns aimed at sensitive data, including crypto wallet information, Gmail accounts, and website registration records.
Bitcoin Red Team found 85 critical flaws in less than 30 hours
Defenders are trying to match that acceleration.
This month, Bitcoin Red Team, a volunteer security group led by Cashu founder and Bitcoin open-source developer Calle and AnchorWatch CEO Rob Hamilton, carried out a large AI-assisted security audit across the Bitcoin open-source ecosystem. The review covered wallets, cryptographic libraries, and infrastructure projects.
In less than 30 hours, the team scanned 390 Bitcoin-related open-source projects and submitted 4,962 security findings. Those included 85 critical vulnerabilities and 635 high-severity issues. The report said the team averaged about 2.31 critical or high-severity findings per person per hour, with scanning costs of about $10,000 per day.
Calle said a large backlog of historical technical debt in open-source code is now colliding with highly efficient AI code-analysis tools. In his view, most unmaintained projects are likely to contain vulnerabilities and should be treated as unsafe by default until proven otherwise.
He singled out Lightning Network software, which is designed to make Bitcoin payments faster and cheaper. Because of its technical complexity, he said its code quality is "worse than average," making audits significantly harder than in other categories.
Calle also argued that the era of complaining about "low-quality PRs" or "low-quality audits" has passed. Projects now need to build their own AI audit pipelines and use AI to quickly filter and reproduce reports. External red-team testing, a security assessment method that simulates real attackers, may need to become a long-running practice. Projects that began building AI-based security and audit processes months ago, he said, are already in a clearer position of strength.
Chinese open-source models are becoming part of the toolkit
The report frames the global AI race as more than a contest over performance and cost. It also describes a split in development models: leading Chinese systems have broadly leaned toward open source, while many top overseas providers have stuck with closed models. That divergence is starting to matter for crypto security.
During the Coldcard attack investigation, Galaxy head of research Alex Thorn said restrictions placed on security research by some U.S. large language models affected researchers’ ability to track stolen funds. Because those tools could not be used smoothly, the team turned to Chinese open-source AI models to help protect user assets and conduct on-chain tracing.
Bitcoin Red Team reported a similar experience. The group said it is currently using Moonshot AI’s Kimi K3 and Zhipu AI’s GLM 5.2 to scan Bitcoin open-source projects for vulnerabilities. Calle said developers routinely run into safety-policy restrictions imposed by model providers such as OpenAI and Anthropic during security research. Even when researchers have completed KYC and applied for Trusted Access, they can still face repeated denials or usage limits.
By contrast, the report says Chinese open-source models come with fewer restrictions, making them more suitable for large-scale code analysis and security auditing. Calle questioned whether current policy settings are constraining white-hat researchers without meaningfully stopping black-hat actors.
BPI and crypto firms want AI labs to open access to defenders
That dispute has now spilled into a broader industry push. A few days ago, the Bitcoin Policy Institute, or BPI, together with dozens of crypto institutions including Anchorage Digital, BitGo, Bitwise, Blockstream, Kraken, Ledger, MARA, and Trezor, released an open letter urging frontier AI labs to establish or expand long-term trusted access programs for Bitcoin and other open-source software developers. The goal is to let security teams use advanced AI earlier to find and fix vulnerabilities.
The letter says frontier AI is rapidly changing the balance between attack and defense in cybersecurity. Advanced models can already analyze large codebases, identify potential vulnerabilities, and accelerate complex technical work, but those same capabilities are also being used by attackers. At the same time, many open-source security teams still lack access to frontier models, while safety restrictions on public AI systems can obstruct legitimate research and push some developers toward weaker open-weight models for critical code review.
The letter specifically says open-source maintainers in crypto, including those working on Bitcoin Core, currently cannot access some AI labs’ cybersecurity access programs. BPI argues that the Bitcoin network now protects more than $1 trillion in assets, and that any vulnerability in open-source infrastructure could put users’ life savings at risk.
BPI also said it has received multiple reports from open-source maintainers who claim sophisticated attackers, including potential foreign adversaries, are using advanced AI capabilities to sustain attacks at a speed small maintenance teams cannot absorb.
In BPI’s view, frontier AI could become one of the most powerful defensive technologies available, but only if defenders protecting critical infrastructure can get those capabilities before attackers fully exploit them. To narrow that gap, BPI called on AI labs to provide certified open-source defenders with early and controlled access to frontier cybersecurity models, including pre-release versions when appropriate; enough compute and long-term agent usage allowances; secure environments for analyzing private or undisclosed code; broader access for small organizations, nonprofits, and independent maintainers; and direct channels with AI lab security teams to coordinate disclosure and remediation.
A new speed race in Bitcoin security
From the Coldcard hack and Boltz’s service suspension to Bitcoin Red Team’s large-scale scanning effort and BPI’s open letter, the report paints a picture of AI moving from a support tool to a central force in crypto security. What is taking shape is a contest over speed, access, and execution.
The side that gets frontier AI earlier, and uses it more effectively for code review, vulnerability discovery, and remediation, is likely to hold the stronger position.