AI Finds $100M Bug for $2, Exposing Critical Self-Custody Risks

Woofun AI reports that Dragonfly managing partner Haseeb Qureshi demonstrated the profound cost-efficiency of artificial intelligence in cybersecurity by using the GLM 5.2 model to rediscover the vulnerability associated with the recent ColdCard theft. The experiment highlighted that the model identified the flaw after researchers directed it toward the affected software, at negligible cost. This outcome serves as a stark illustration of how attackers and developers can now examine code protecting large amounts of cryptocurrency at negligible expense, fundamentally altering the economic landscape of digital asset security. There is no evidence that the original thieves utilized AI, but the test confirms that the barrier to entry for serious vulnerability research has collapsed, allowing for rapid, low-cost scrutiny of critical infrastructure.

The financial implications of this shift are profound, particularly when considering the scale of assets at risk. Qureshi’s experiment revealed that a mere $2 expenditure could uncover a weakness linked to the recent ColdCard theft, a disparity that redefines the risk-reward calculus for both defenders and adversaries. Traditionally, serious vulnerability research demanded specialist knowledge and countless hours of manual labor, creating a high barrier to entry that limited the scope of code review.

Now, the ability to inspect code protecting millions of dollars for a few dollars per attempt means that attackers can afford to conduct thousands of unsuccessful attempts without significant financial loss. This economic logic suggests that even if most searches produce nothing useful, the cost of failure is so low that the potential reward of finding one valid flaw in software protecting millions of dollars justifies the volume of trials. Consequently, the sheer scale of Bitcoin’s value becomes a magnet for automated, low-cost probing rather than a deterrent.

Structurally, the industry is witnessing a decisive shift from manual to automated code review processes. An experienced human researcher can only inspect a limited number of firmware releases, smart contracts, and cryptographic libraries within a reasonable timeframe, creating inherent bottlenecks in security auditing. AI allows far more code to receive an initial review by enabling researchers to assign several models to the same codebase simultaneously.

These models can be instructed to investigate different failure paths, and the process can be repeated whenever a stronger model becomes available, creating a continuous loop of improvement and scrutiny. This automation does not replace the need for security expertise but rather lowers the cost of generating leads, testing theories, and covering more code than a human researcher could inspect alone. The result is a broader, faster coverage of the codebase, where the initial filter is applied by machines, allowing human experts to focus on verification and complex analysis.

The economic logic of vulnerability hunting has been inverted by these technological advancements. In the past, the high cost of manual research meant that only the most promising targets were scrutinized deeply, leaving many areas of code unexamined due to resource constraints. Now, someone looking for weaknesses can examine many targets without needing every run to succeed, as the cost of an individual attempt is minimal.

This democratization of security testing means that attackers can cast a wider net, scanning old firmware and automating repetitive analysis to test possible weaknesses across multiple platforms. The ability to conduct thousands of unsuccessful attempts for a fraction of the previous cost means that the probability of finding a critical flaw increases significantly over time. For developers, this implies that legacy code, which may have been considered secure due to lack of scrutiny, is now vulnerable to exhaustive, low-cost automated reviews that were previously economically unfeasible.

Per Woofun AI, the experimental constraints and timeline of the GLM 5.2 test provide critical context for understanding its significance. Qureshi stated that researchers asked several models to examine the software after the ColdCard incident had already focused attention on it, meaning the models were not searching blindly across every hardware wallet. The models were not told where the vulnerability was located, yet GLM 5.2 found it in roughly 20 minutes, demonstrating the speed at which modern AI can process and analyze complex codebases.

This rapid identification occurred because the context was already established; researchers knew that the relevant software contained a serious problem, providing a starting point for the AI’s investigation. The model was not discovering an attack from scratch but rather rediscovering an existing weakness with remarkable efficiency. This timeline underscores the potential for AI to accelerate the discovery phase of security audits, reducing the window of opportunity for attackers to exploit known but unpatched vulnerabilities.

It is crucial to distinguish between finding bugs and exploiting them at scale, a nuance often overlooked in discussions about AI’s impact on security. Identifying a bug is different from exploiting it, as an attacker may still need to perform additional steps to convert a theoretical weakness into a practical theft. Qureshi’s comments do not show that GLM 5.2 performed those steps or that an AI system participated in the original theft, but they do establish that a general-purpose model can rediscover the underlying technical weakness quickly and at negligible cost.

The experiment highlights the AI’s capability in the reconnaissance and analysis phases, rather than the execution phase of an attack. This distinction is important because it suggests that while AI can lower the barrier to finding vulnerabilities, human expertise remains essential for understanding how to exploit them or, conversely, how to patch them effectively. The AI serves as a powerful tool for lead generation, but the final steps of an attack or a fix still require nuanced understanding and precise execution.

The debate around long-term security threats to Bitcoin often focuses on quantum computing, which presents a different set of challenges compared to AI-assisted code analysis. A sufficiently capable quantum machine could eventually threaten the signature schemes protecting coins, forcing networks and users to migrate to new cryptography, a risk that requires advance preparation and significant infrastructure changes. In contrast, AI-assisted code analysis presents an immediate problem because the tools can already be used against software running today, without waiting for technological breakthroughs in quantum mechanics.

AI does not need to defeat Bitcoin’s underlying mathematics; it can search the surrounding code for mistakes such as weak random-number generation, exposed secrets, faulty cryptographic implementations, and unsafe recovery functions. This immediacy means that the threat from AI is not hypothetical but current, affecting wallets, bridges, libraries, and applications that are actively protecting user funds right now. The focus must therefore shift from preparing for a distant quantum threat to addressing the present-day vulnerabilities exposed by AI-driven code analysis.

Specific software vulnerabilities targeted by AI include those that stem from human error rather than fundamental cryptographic flaws. AI can be directed at wallets, bridges, libraries, and applications currently protecting user funds, scanning for mistakes such as weak random-number generation, exposed secrets, faulty cryptographic implementations, and unsafe recovery functions. These errors are often subtle and difficult for human auditors to catch consistently, especially when reviewing large volumes of code.

The immediate concern is therefore a human mistake in the software around Bitcoin rather than AI breaking Bitcoin itself, as the AI is merely amplifying the visibility of existing weaknesses. An attacker can use a model to scan old firmware, automate repetitive analysis, and test possible weaknesses, while a wallet manufacturer can use the same model before release, directing it toward key generation, transaction signing, and recovery procedures. This dual-use nature of AI means that both attackers and defenders have access to the same powerful tools, creating an arms race where the speed and accuracy of code analysis determine the outcome.

The randomness flaw identified in the ColdCard incident highlights the specific risks to self-custody wallets, which are often perceived as the most secure option for storing Bitcoin. Qureshi described the affected users as people who had followed the usual security advice, moving Bitcoin away from exchanges and storing it in dedicated hardware wallets, yet they were still vulnerable due to a flaw in the device’s firmware. The affected firmware used a weaker source of randomness than intended, meaning that if a private key is generated from predictable inputs, keeping the device offline cannot make the key stronger afterward.

An air gap can reduce exposure to remote attacks against a properly generated secret, but it cannot repair a secret that was weak from the moment it was created. Hardware-wallet users therefore rely on more than their own handling of a recovery phrase; they also trust the manufacturer’s firmware, hardware design, cryptographic libraries, and random-number generator. This dependency on the manufacturer’s code quality means that self-custody is only as secure as the software that generates and protects the keys, a fact that is often overlooked by users who assume that physical isolation guarantees security.

Future security protocols must adapt to this new reality, where AI-assisted analysis becomes a standard part of development rather than a post-mortem tool. Wallet manufacturers and crypto protocols should rescan older releases when major models improve, especially when the code creates private keys, signs transactions, or controls wallet recovery, as software audited five years ago can now be examined with tools that were unavailable when it reached customers. A firmware update will not solve every problem, as correcting the software may protect new wallets while leaving older ones exposed, requiring users to generate new keys under the repaired system and transfer their funds.

Providers should therefore be able to identify affected versions, contact customers quickly, and explain how to migrate safely, a process that also creates opportunities for scammers who exploit the confusion. Users should never provide a seed phrase or private key to a support representative, website, social-media account, or AI chatbot, as a legitimate recovery process does not require another person or service to receive those secrets. This marks a critical shift in user education, where the threat landscape now includes sophisticated AI-driven scams that target users during security incidents, necessitating heightened vigilance and clear communication from providers.