Over 95% of Coinbase's code is now written by artificial intelligence. The company that holds billions in user assets is increasingly a product of algorithms. This isn't innovation—it's an experiment on an industrial scale.
Context: The Efficiency Mirage
In July 2025, Coinbase's Platform Lead, Rob Witoff, dropped a bombshell: 95 to 100 percent of the code at the exchange is now AI-generated or AI-assisted. This is a dramatic leap from the 40% figure disclosed just five months earlier in February. Simultaneously, the company announced a 14% workforce cut, laying off 700 employees. The narrative is clear: AI replaces humans, boosts output, and slashes costs. But as a crypto security audit partner who has spent over a decade dissecting smart contracts and exchange architectures, I see something else—a structural transformation that trades long-term accountability for short-term productivity.
Coinbase is not a startup experimenting with AI on the side. It is a publicly traded, SEC-regulated custodian of digital assets. Its infrastructure handles custody, trading, staking, and the Base Layer-2 chain. The claim that AI now generates nearly all its code is not just a PR statement; it is a auditable claim exposed to investor scrutiny. Yet the market has largely celebrated this as a breakthrough, ignoring the forensic reality: code generated by large language models (LLMs) carries a unique fingerprint of statistical approximation, not deterministic logic.
Core: The Autopsy of AI-Generated Code
Let me start with what I know from firsthand experience. In my audits of smart contracts, I have seen how even human-written code fails under edge cases. Now imagine that 95% of that code was produced by an AI that learned from public GitHub repositories, Stack Overflow snippets, and documentation—including buggy examples. The risk is not theoretical; it is structural.
Coinbase has implemented a critical safety boundary: core cryptography remains human-reviewed. This is smart, but incomplete. The exchange's trading engine, risk management, user balance tracking, and fee calculation logic are all likely AI-generated. An error in any of these modules could lead to exploit losses that dwarf the estimated $4 million I helped save during the DeFi Summer liquidity drain investigation in 2020. In code, silence is the loudest vulnerability. If an AI introduces a subtle integer overflow in the matching engine, the first sign may be a billion-dollar arbitrage flash crash.
Consider the numbers: AI agents now perform the equivalent work of 1,200 full-time employees. Each engineer manages five to ten AI agents. This ratio is unprecedented in financial technology. It means that the human oversight bandwidth is stretched thin. An agent can execute hundreds of code commits per day. A human can review maybe ten. The audit trail becomes opaque: who is responsible for a bug introduced by an AI agent that was prompted by an engineer who didn't fully understand the model's training data? Standardization fails when it ignores human chaos.
Moreover, the speed of this transition is alarming. Going from 40% to 95% AI code in five months implies a complete rewrite of vast swathes of the codebase. That is not iterative improvement; it is a forklift upgrade. The risk of regression bugs—where previously working features break—is high. The company's own data shows that 100% of internal prototype projects are now AI-automated. This means that new product development is entirely outsourced to models that, as of today, still hallucinate plausible but incorrect code.
The Forensic Narrative Accountability
A key part of my work involves tracing blame after an incident. When I dissected the Terra/Luna collapse, I could point to specific smart contract functions that failed. With AI-generated code, attribution becomes fuzzy. The model's training data may include patterns from projects that were themselves compromised. The code may work under normal conditions but fail catastrophically under stress—just like Terra's algorithmic stablecoin. The blockchain remembers, but the auditors forget. If Coinbase suffers a major exploit, the forensic analysis will hit a wall: the code was written by an algorithm that the engineers themselves cannot fully explain.
Coinbase's layoffs compound this. Firing 700 employees, many of whom could have been the ones who understood the previous codebase organicly, creates institutional memory loss. The new hybrid team of fewer humans and many AI agents lacks redundancy. If an agent fails, who steps in? The remaining engineers are now managers of AI, not developers. Their skills atrophy as they shift from writing code to reviewing it. This is a long-term talent risk.
Contrarian: What the Bulls Got Right
I am not here to dismiss the potential upside. The efficiency gains are real: AI agents doing the work of 1,200 employees means Coinbase's operating costs will drop significantly. This could translate into lower trading fees, faster feature deployment, and better margins. The market has priced in some of this, but the sheer scale—predicting 100,000 AI employees equivalent by 2030—creates a narrative multiplier. For a stock like COIN, this is gold.
Investors are betting that Coinbase's first-mover advantage in AI adoption will create a moat. They argue that competitors like Binance or OKX will struggle to replicate such deep integration because of their own regulatory constraints or infrastructure debt. There is truth here: Coinbase has always been more transparent and compliant, and its public disclosure of AI code percentages is a sign of confidence.
But the contrarian angle is more subtle. The real value driver may not be cost reduction but the ability to spin up new products rapidly. The Base chain could benefit from AI-accelerated smart contract development, auditing, and tooling. If Coinbase can deploy standard auditing or security tool for Base ecosystem projects using AI, it could attract developers. Yet this assumes the AI-generated audit tools are themselves reliable. That is a circular dependency.
The Unspoken Risk: Trust Degradation
What the bulls miss is that cryptonetworks are built on trust—not just in code, but in the human institutions that govern that code. Coinbase has historically differentiated itself through human integrity: its leadership, compliance team, and engineering culture. By replacing human judgment with AI in the software production line, it risks losing the very trust that made it the preferred on-ramp for institutional capital.
Consider the message to regulators: "We can't fully explain how our code works because an AI wrote it." The SEC, already aggressive in crypto enforcement, will not look kindly on a black-box exchange. The recent series of enforcement actions against exchanges for market manipulation and system failures could easily extend to AI-generated vulnerabilities. Coinbase's decision to keep core crypto human-reviewed is a nod to this, but it raises the question: why trust AI with everything else?
Takeaway: The Accountability Gap
The real question from my cold dissector perspective is not whether AI can write code. It can. The question is whether we, as an industry, are willing to accept code that is not fully understood by any human. In my experience auditing over 100 protocols, the most dangerous bugs are not the ones that are complex; they are the ones that seem correct to both the developer and the reviewer. AI magnifies this: it generates code that looks plausible but may contain hidden invariants that violate economic security.
Coinbase is making a bet that the efficiency gains outweigh the risks. For now, the market agrees. But as a partner who has seen both the promise and peril of rapid tech adoption, I advise caution. Liquidity is a mirror, not a vault. It reflects the trust of its users. If that mirror cracks—because of an AI-induced incident—the exodus will be swift and brutal. The blockchain will remember the transaction, but the humans who trusted Coinbase will remember the failure.
In conclusion, Coinbase's AI transformation is a case study in the tension between speed and safety. The exploit wasn't in the code; it was in the assumption that we could scale human oversight through automation. I have seen this pattern before: standardizing chaos without understanding the messy reality of human error. Standardization fails when it ignores human chaos. The market may celebrate today, but the true test comes the first time an AI-generated bug triggers a withdrawal freeze. Let's hope the humans still know how to flip the kill switch.
