The Ghost in the Machine Has Teeth: FSB Warns AI-Driven Attacks Are Now Systemic

The warning arrived like an anonymous leak from a quiet room in Basel. The Financial Stability Board—the body forged from the ashes of the 2008 collapse, designed to peer into the darkest corners of global finance—issued a statement that felt less like a recommendation and more like a confession. Artificial intelligence, it said, is no longer just a tool for optimizing portfolios or flagging fraud. It is a weapon. And it is about to become a systemic threat to global financial stability. I read the FSB statement on a muted Monday, coffee growing cold, while a torrent of on-chain data flickered across a secondary screen. The words were careful, bureaucratic, almost anaesthetized. But beneath the formal syntax lay a quiet ruin: the machines that we’ve built to run the world’s money—the payment rails, the clearing systems, the oracle-like trust layers—are now vulnerable to other machines that have learned to attack. Not via brute force, but via evolution. The ghost in the machine has learned to meet the market, not just to whisper to it—but to bite. Let me pull back the curtain for a moment. The FSB is not a random think tank. It was born from the wreckage of Lehman Brothers in 2009, specifically to identify fragilities before they cascade into global panic. When it speaks, central banks and treasury officials at least pretend to listen. But this particular warning, issued in the shadow of a brutal bear market for crypto and a confusing AI boom, is unlike anything we’ve seen. It doesn’t cite a specific incident. It doesn’t name a victim. Instead, it simply states the obvious that no one wanted to say out loud: AI-driven cyber attacks have evolved from a theoretical nuisance into a systemic risk that could break the financial system. Not a bank. Not a trading venue. The entire system. I’ve spent the last decade staring at code during the day and lying awake thinking about trust during the night. In 2017, I spent six months auditing Uniswap’s constant product formula in a tiny Buenos Aires apartment, trying to understand how an algorithm could replace a market maker. I wrote a piece called “Liquidity as Trust,” and I was right about that. But what I didn’t fully grasp was that trust and vulnerability are two sides of the same coin. The more we automate trust, the more power we hand to the algorithm. And algorithms, as the FSB now concedes, can be turned against us. They can be taught to find the cracks in our new cathedrals of finance. Consider the attack surface. The global financial system is a dense, hyperconnected web of high-value targets. Core trading engines, payment and settlement systems, and the SWIFT messaging network that carries trillions of dollars daily—these are not fortress-like structures. They are legacy software patches over even older code. And now, AI-powered attackers are using reinforcement learning to autonomously discover attack paths that a human hacker would take days to map. They’re using generative AI to generate phishing emails that no longer look like phishing—they carry the vocabulary, the cadence, and the emotional undertones of a specific CFO, constructed from just a few public earnings calls. They are using adversarial examples to trick malware detectors into seeing a clean file where a log4j-style exploit sleeps. The FSB’s warning is not about a random hack. It’s about a paradigm shift in offensive cyber capabilities. Traditional attacks were rule-driven: a human writes code, points it at targets, and hopes the exploit works. AI-driven attacks are completely different. They are learning-driven and adaptive. They observe the environment, test hypotheses, and iterate their strategy in real time. We’ve crossed a threshold where the AI doesn’t just speed up human tactics—it invents new ones. The code remembers what the market forgets, and the market is about to remember a very painful lesson. Reading the silence between the blocks, I see something the mainstream financial press misses. The FSB’s warning wasn’t an accident. It was a deliberate signal, carefully calibrated to reach institutions that still operate as if a firewall is enough. But the deeper truth is that the FSB’s internal assessments, which we don’t see, must contain specific case studies or red-team exercises that scared them enough to act publicly. The warning acknowledges, implicitly, that existing security frameworks—the ones that rely on firewalls, intrusion detection, and signature-based antivirus — are insufficient in the age of adaptive adversaries. If they were not, no such warning would be needed. The quiet admission is that our defenses have been operating in an era of ancient warfare, while the attackers have already discovered gunpowder. Let me be more technical, because this matters. In the world of AI security, we talk about three attack vectors. The first is AI-powered social engineering, where generative models craft highly convincing messages at scale. A bank in Hong Kong reportedly lost millions after a deepfake video call convinced a branch manager to authorize transfers. That was a preview, not a novelty. The second is autonomous vulnerability discovery. Reinforcement learning agents can scan millions of lines of code, probing for weaknesses with a persistence no human team could match. They can find the one misconfigured microservice in a sprawling enterprise architecture, the one forgotten API that connects to a settlement engine. The third is adversarial machine learning itself: poisoning data sets used by fraud detection models, injecting subtle inputs that cause an AI system to misbehave in dangerous ways. Imagine an AI that controls a trading algorithm and is slowly manipulated through adversarial signals to position the portfolio in exactly the wrong way, right before a market event. That is not science fiction. That is the logical consequence of automated decision-making without resilient design. And here is where the FSB gets to its most consequential recommendation: the call for “diversifying technology dependencies.” That phrase sounds textbook, but it is a Hail Mary. It means the financial system has become dangerously dependent on a small number of technology providers. Everyone uses the same cloud providers, the same security vendors, the same database platforms. When an AI attack tears through yesterday’s endpoint detection, it doesn’t just break one bank; it breaks every bank using that common stack. The monoculture is systemic. In crypto, we call this the “walled garden” problem, but in traditional finance it’s called interoperability risk. The FSB knows that diversification is a blunt instrument—you can’t just tell every bank in the world to build its own Google Cloud. But the recognition is a cry in the dark: our complexity has created a single point of failure, and that failure point is now openly targetable. Now, let me speak honestly as someone who watched Terra’s algorithmic stablecoin die in 2022, recording every silence from a cabin in Patagonia. I saw math fail not because the code was wrong, but because the incentives were flawed. The FSB warning is different. It’s not about incentives; it’s about malicious intelligence. But there is an eerie parallel. We tend to trust that an emergency will unfold slowly enough for decisions to be made. That is no longer true with AI-driven attacks. An AI attacker does not need to sleep. It doesn’t take weekends. It doesn’t panic when the market drops. It simply continues to probe, adapt, and exploit. The asymmetry is brutal—defenders must cover every possible attack surface, while attackers only need to find one gap, and they can use AI to find that gap faster than any human. What does this mean for crypto? The FSB report was published on a crypto media site, and I suspect that is not a coincidence. Crypto exchanges and DeFi protocols are increasingly part of the global financial fabric. They are also, in many cases, less protected than their regulated counterparts. Many have insufficient bug bounty programs, no formal red-team exercises, and a dangerous over-reliance on open-source code that has never been truly audited. In 2024, I collaborated with a group of traditional finance experts on the BlackRock Bitcoin ETF filing, and we saw a gap in perspective. Traditional finance has heavy compliance overhead—some of it useless—but it has robust incident response frameworks. Crypto has speed, radical transparency, and a community of hackers who test systems around the clock. The FSB warning implicitly signals that both worlds face a common enemy. AI attacks don’t care if the target is a bank in Frankfurt or a smart contract in a decentralized exchange. They only care about value concentration and weak signals. But here is the contrarian angle that makes me uneasy. The FSB’s warning, while necessary, is also convenient. It gives central banks and established financial institutions a new rationale to tighten control over the digital economy. “AI safety” is a malleable concept. It can be a genuine commitment to building robust systems—or it can be a tool for surveillance, a reason to require “AI audits” that are actually backdoor access. I have seen the seed of this in the MiCA regulation, where every stablecoin reserve is now audited to the point of strangling small projects. The warning may end up being the justification for a new wave of compliance requirements that only the largest institutions can afford. The phrase “diversifying technology dependencies” could be twisted into a mandate to use only approved cloud providers—concentrating power even further. Let me also flag the insurance dimension. The FSB warning will inevitably force cyber insurers to repricing their portfolios. AI-related attacks are a black box. You cannot model the probability of an intelligent attacker that adapts to your defenses. The result will be one of two extremes: AI attacks will be explicitly excluded from cyber insurance policies, leaving institutions entirely exposed, or premiums will skyrocket to a level where only mega-banks can afford full coverage. Either way, the cost of demonstrating “AI security” will fall hardest on smaller firms—community banks, credit unions, and, in crypto, a host of decent protocols that might be forced to buy compliance theater no one fully understands. The code remembers what the market forgets, and the market seems to have forgotten that we are not just fighting a technology problem. We are fighting an intelligence problem. We cannot simply buy another firewall. We need to build systems that are fundamentally different in their topology—less centralized in their architecture, less reliant on common libraries, and more transparent about their failure modes. In the crypto world, we have an advantage: we can encode automated games, economic incentives, and governance mechanisms to create resilience. But we have not done it yet. Most DeFi projects are still piles of smart contract logic waiting for a pathologist. When the herd wakes, the signal has already faded. We need to act before the first major AI-driven financial crisis goes viral. The FSB warning is that signal, and markets are not yet pricing the risk. I am not talking about volatility pricing, but rather about the strategic value of companies that provide AI security for finance. The trend is unmistakable: regulatory pressure will create demand for AI audits, red-teaming, and adaptive defense platforms. We will see a bifurcation of the cybersecurity market into legacy players and AI-native defenders. The latter will become the new guard of the industry. But investors beware: many “AI security” startups are just wrappers around an LLM and a marketing deck. The true signal lies in firms that demonstrate an integrated approach to threat detection, autonomous response, and—most crucially—a commitment to explainability. Let me close with a question, because that is what we do when we are afraid. If an AI can learn to lie convincingly as our CFO, and to search for vulnerabilities in our trading infrastructure while we sleep, what is the role of human wisdom in finance? We traded chaos for consensus over the last few decades, outsourcing risk modeling to algorithms that cannot understand context. Now those algorithms are being turned against us. The FSB’s warning is a moment to pause. Not to retreat from technology, but to remember that trust is a human artifact. No smart contract can be trusted to be its own guardian. No offering document can prepare us for an adversary that learns faster than we can regulate. We need to build a new system of checks and balances—not just in code, but in governance, in diversity of thought, and in our willingness to admit that we have never been as safe as we thought. The quiet ruin when the algorithm broke will be a collapse of trust before it is a collapse of the network. The FSB just told us to prepare. Are we listening?