In early July 2026, an event rippled through the Ethereum research community that many dismissed as a mere parlor trick. Franklyn Wang, a little-known AI researcher, publicly identified Vitalik Buterin as the anonymous editor behind a key revision to EIP-7503 — a privacy proposal that enables zero-knowledge wormhole transfers. Wang’s method wasn’t based on IP tracking, metadata analysis, or even stylistic word choice. He used an AI model, Co-Invest, to analyze what he calls the “thought fingerprint” — the unique logical structure through which a person explains mathematical algorithms. And it worked.
This is not a story about AI surveillance. It is a story about the collapse of the assumption that anonymity in open-source contributions can be preserved through simple technical precautions. For anyone who has ever believed that code speaks for itself, this experiment suggests that the mind behind the code speaks louder.
Context: The EIP-7503 Anonymity Challenge
EIP-7503, proposed by Keyvan Kambakhsh, aims to bring “wormhole privacy” to Ethereum using zero-knowledge proofs. It allows users to send messages without revealing the sender’s identity. In line with Ethereum’s open ethos, Kambakhsh allowed anonymous edits to the draft. At some point, a contributor using a disposable account made substantial technical revisions, fixing a critical flaw in the zero-knowledge circuit design. No one questioned the identity of the editor — until Wang decided to test a hypothesis.
Wang fed the entire revision history into Co-Invest, an AI research engine that can cross-reference documents and infer authorship patterns. The model flagged the anonymous editor’s reasoning style — specifically, its reliance on algebraic rather than geometric intuition, its preference for modular proofs over linear ones, and a characteristic way of explaining edge cases — as statistically anomalous. Among 10 million Ethereum developers, Wang’s model assigned the highest probability to Vitalik Buterin, though with only 20% confidence. But that was still 10 times higher than the next candidate. The AI had traced a logical pattern back to a single individual, even when all traditional identifiers were stripped away.
Core Analysis: The Mechanics of Cognitive De-Anonymization
Dissecting the atomicity of the reasoning structure. Traditional stylometry analyzes word frequency, sentence length, and punctuation habits. Wang’s approach goes deeper. It examines the sequence of cognitive steps: how a writer decomposes a problem, which mathematical frameworks they default to, and how they handle abstraction layers. In Buterin’s case, the model detected his signature habit of first laying out the simplest failure case, then generalizing gradually — a pattern visible in his early Ethereum Yellow Paper drafts and his blog posts on sharding.
Why does this matter for blockchain privacy? Because countless protocols assume that anonymity is preserved as long as you don’t use a known address or IP. *This experiment proves that the way you think is a biometric identifier as unique as a fingerprint, and it can be extracted from any sufficiently long piece of technical writing.*
I have spent years auditing Layer 2 scaling solutions, from state channel race conditions to ZK-rollup circuit constraints. Tracing the gas limits back to the genesis block is a familiar exercise — but tracing a thought process back to a specific developer is a new kind of forensics. In 2017, I identified race conditions in Raiden Network’s settlement logic by reading the same reasoning patterns I had seen in earlier Vitalik posts. Back then, it was a hunch. Now it’s an AI model.
The layer two bridge is just a pessimistic oracle — it has to assume the worst about the connected chain. Similarly, any anonymity layer that assumes “different vocabulary” suffices to hide identity is now a broken oracle. The pessimistic assumption must be: your cognitive style is public, and it can be matched.
Contrarian Angle: The 20% Confidence Trap
It is tempting to react with panic. A 20% confidence score seems low — far from legal proof. But that misses the point. The model’s ranking ability, not its absolute certainty, is the weapon. In a targeted investigation, a 10x higher score than any other candidate is enough to justify further probing. Three independent models each with 20% confidence would compound to near certainty. The technology is immature, but the signal is already actionable.
The contrarian risk here is that the crypto community overreacts in one of two ways: either dismissing it as a “novelty” or rushing to implement ill-thought-out defenses like uniform writing templates. The first is naive (the pattern is real), the second is futile (templates can be reverse-engineered). The more subtle danger is that privacy advocates abandon textual contributions altogether, moving to voice or video, which introduce even more identifying features.
Finding the edge case in the consensus mechanism is hard work — but finding the edge case in human cognition is even harder. The edge case here is that Wang’s model worked only because Buterin’s writing is exceptionally idiosyncratic. Most developers do not have such a distinct mathematical signature. But the ones who do — the core researchers, the protocol architects — are precisely the individuals whose anonymity matters most. If they can be identified, the security of future anonymous proposals is compromised.
Takeaway: From Linguistic Privacy to Cognitive Privacy
This single experiment will not de-anonymize the entire Ethereum developer base. But it marks a shift in the threat model for blockchain privacy. The era of assuming that “code is not a personality” is over. The next generation of privacy tools must obfuscate not just the words we write, but the logical paths we walk.
What happens when an AI model can identify the author of a DAO proposal, a key upgrade rationale, or a governance argument? Will regulators use thought fingerprints to track down anonymous token holders? Will project founders be required to submit to “cognitive KYC” before being allowed to contribute? These questions are no longer theoretical.
Mapping the metadata leak in the smart contract was yesterday’s problem. Today, the metadata leak is in our minds.