AI’s Dual Edge: Revolutionizing Crypto Security and Cyber Warfare

Imagine the unsettling scenario: your precious crypto tokens, diligently secured in a hardware wallet you believed to be an impenetrable fortress, are suddenly compromised. Within minutes, hackers, leveraging the relentless efficiency of artificial intelligence, uncover the “master key” to your digital vault.

This is no longer a distant threat. Large AI models are rapidly advancing into the core of cryptocurrency security, simultaneously enhancing both offensive capabilities and defensive strategies.

As AI models continuously evolve, they are, on one hand, dramatically lowering the technical barriers for complex crypto attacks, accelerating vulnerability exploitation and the speed of breaches. On the other, they are paradoxically emerging as indispensable weapons for the crypto industry to identify weaknesses, pinpoint risks, and expedite critical repairs. This transformative trend is particularly pronounced within the Bitcoin ecosystem, a behemoth boasting a market capitalization exceeding one trillion US dollars.

In the global race for AI supremacy, Chinese domestic large models, capitalizing on their robust capabilities and open-source ecosystem advantages, are increasingly stepping up as “first responders” for security audits within the crypto community.

The AI-Driven Surge in Cyberattacks

The crypto world was recently rocked by a massive hack targeting the hardware wallet Coldcard. Over $100 million in Bitcoin assets were reportedly pilfered, a staggering loss that not only depleted funds but also significantly eroded market confidence in Bitcoin’s self-custody security.

Coldcard has long been lauded as a high-security “safe.” Given its open-source codebase, the community speculates that attackers likely employed AI to scrutinize older firmware versions, unearthing a vulnerability that had remained hidden for five years within vast historical code. Compounding concerns, post-attack analysis revealed that the AI model Claude Code could precisely locate this vulnerability in a mere eight minutes.

Further illustrating this escalating threat, non-custodial Bitcoin swap service Boltz announced an indefinite suspension of its services this month. A key reason cited was that the relentless pace of AI-assisted attacks had outstripped the team’s ability to patch vulnerabilities. Boltz had endured continuous automated AI probing attacks and addressed multiple incidents over several months. However, the recent acceleration in attack frequency, coupled with suspicions of multiple well-resourced attack groups simultaneously targeting their platform, made secure operation during remediation impossible. Ultimately, Boltz suspended its swap services, which were later taken over by an anonymous Bitcoin team.

As large AI models become more sophisticated, hackers are weaponizing them, drastically reducing the cost and increasing the speed of crypto attacks. Historically, discovering complex vulnerabilities demanded extensive time for manual code review and meticulous attack path design. Today, AI is automating an increasing number of these processes, enabling attackers to analyze code, pinpoint vulnerabilities, and even orchestrate more clandestine exploits and attack workflows with unprecedented efficiency and lower overhead.

For example, the North Korean hacking collective Kimsuky has recently broadened its use of generative AI. Beyond crafting highly effective phishing lures, the group has established an independently operating local large language model (LLM) environment and a retrieval-augmented generation (RAG) system. Their objective is to further automate intelligence extraction and streamline attack processes. In their specific attack methodologies, Kimsuky now leverages generative AI to produce highly convincing virtual asset and financial sector documents, facilitating more targeted spear-phishing campaigns aimed at sensitive data such as crypto wallet information, Gmail accounts, and website registration records.

Proactive Defense: Projects Must Build Their Own AI Firewalls

With offensive capabilities rapidly advancing, defensive strategies must accelerate to keep pace.

This month, the Bitcoin Red Team, a volunteer security organization spearheaded by Cashu founder and Bitcoin open-source developer Calle, alongside AnchorWatch CEO Rob Hamilton, undertook a comprehensive AI-assisted security audit of the Bitcoin open-source ecosystem. This extensive review encompassed critical projects including wallets, cryptographic libraries, and foundational infrastructure.

In less than 30 hours, the team meticulously scanned 390 Bitcoin-related open-source projects, unearthing a staggering 4,962 security findings. This included 85 critical vulnerabilities and 635 high-risk vulnerabilities. On average, each team member identified approximately 2.31 critical or high-risk issues per hour, with the daily scanning operation costing an estimated $10,000.

Calle highlighted a critical observation: the vast accumulation of historical technical debt within open-source codebases is now directly colliding with the formidable capabilities of high-performance AI code analysis tools. He asserted that the majority of unmaintained projects are highly susceptible to vulnerabilities and should be presumed insecure until proven otherwise. The Lightning Network software, designed for faster, more cost-effective Bitcoin payments, was singled out due to its inherent technical complexity, presenting a “worse than average” code state and significantly higher auditing challenges compared to other categories.

Calle emphasized that in this new AI era, the days of lamenting “low-quality pull requests” or “subpar audits” are over. Projects are now compelled to establish their own AI-driven audit pipelines to swiftly filter and reproduce security reports. Moving forward, external red team testing – security assessments that simulate real-world attacker methodologies – may need to become a continuous, long-term endeavor. Projects that proactively began constructing their AI security and audit processes months in advance are already positioned with a distinct advantage.

Chinese Open-Source Models Fill the Gap, Industry Calls for AI Lab Access

In the global competition for large AI models, Chinese and international vendors are not only vying for superior performance and cost-efficiency but are also charting divergent development paths. While leading Chinese models largely embrace an open-source philosophy, most top international players maintain closed-source strategies.

This fundamental difference in approach is now significantly impacting the crypto ecosystem’s security capabilities. Unexpectedly, in contrast to their closed-source Western counterparts, Chinese open-source models are emerging as vital tools for the Bitcoin community in critical areas such as code review, vulnerability discovery, and security research.

During the investigation into the Coldcard attack, Alex Thorn, Director of Research at Galaxy, revealed that restrictions imposed by some US large language models on security research hindered investigators’ ability to trace stolen funds. Unable to effectively utilize these models, the team was compelled to turn to Chinese open-source AI models to assist in protecting user assets and conducting on-chain tracking.

A similar scenario unfolded during the Bitcoin Red Team’s security audit. The team disclosed their reliance on Chinese AI models, specifically Kimi K3 from Moonshot AI and GLM 5.2 from Zhipu AI, for vulnerability scanning of Bitcoin open-source projects. Calle noted that during security research, developers frequently encounter limitations imposed by model providers like OpenAI and Anthropic. Even after completing KYC and applying for Trusted Access, researchers often face rejections or usage restrictions. In stark contrast, Chinese open-source models present fewer such impediments, making them more suitable for extensive code analysis and security audits. This situation led Calle to question the efficacy of current policies, which appear to constrain white-hat researchers while failing to adequately curb malicious black-hat activities.

Just days ago, the Bitcoin Policy Institute (BPI), in collaboration with dozens of prominent crypto institutions including Anchorage Digital, BitGo, Bitwise, Blockstream, Kraken, Ledger, MARA, and Trezor, issued a powerful open letter. The letter urgently called upon leading AI labs to establish or expand long-term trusted access programs for Bitcoin and other open-source software developers, thereby empowering security teams to leverage advanced AI for proactive vulnerability remediation.

The open letter underscored that cutting-edge AI is rapidly reshaping the cybersecurity offense-defense landscape. While advanced models are demonstrating remarkable capabilities in analyzing vast codebases, identifying potential vulnerabilities, and accelerating complex technical tasks, these very same powers are regrettably being harnessed by attackers. Concurrently, many open-source security teams currently lack adequate access to these frontier models. Furthermore, the inherent security restrictions of public AI models can inadvertently impede legitimate security research, compelling some developers to rely on less capable open-weight models for critical code review, potentially compromising security.

The letter specifically highlighted that open-source crypto maintainers, including those responsible for Bitcoin Core, are presently unable to access cybersecurity programs offered by certain AI labs. Given that the Bitcoin network currently safeguards over $1 trillion in assets, any vulnerability within its open-source infrastructure poses a grave threat to users’ life savings. The BPI further revealed that it has received multiple reports from open-source maintainers detailing how sophisticated attackers, including potential foreign adversaries, are leveraging advanced AI capabilities to launch relentless attacks at a pace that small maintenance teams find unsustainable.

BPI emphasized that frontier AI holds immense promise as one of the most potent defensive technologies available. However, this potential can only be realized if the defenders protecting critical infrastructure gain access to these capabilities before attackers exploit them. To bridge the widening offense-defense capability gap, BPI urged AI labs to: provide certified open-source defenders with early and controlled access to frontier cybersecurity models, including pre-release versions where appropriate; offer sufficient computing power and long-term agent usage quotas; establish secure environments for analyzing private or unreleased code; extend access to small organizations, non-profits, and independent maintainers; and foster robust collaboration with AI labs.

In conclusion, AI is propelling crypto security into an unprecedented race of speed and capability. Moving forward, those who can swiftly acquire and effectively deploy frontier AI will undoubtedly hold a significant advantage in this evolving offense-defense dynamic.