AI Did Not Break ECDSA — the Public Keys Were Already Exposed

Justin Drake typed two words into a research thread and the timeline caught fire: bunker mode. The Ethereum Foundation researcher warned that rapid AI progress could expose weaknesses in ECDSA, the signature scheme securing Bitcoin, Ethereum, and almost every chain that matters. Within hours the phrase had been recycled into a dozen threads. AI is coming for your private keys. The math says otherwise. I spent the morning re-reading the secp256k1 parameters and the original threat model. The claim is not just overstated. It is technically inverted. ECDSA does not rest on pattern recognition. It rests on the elliptic curve discrete logarithm problem — a precise hardness assumption about integer arithmetic. Classical AI, the kind that finds statistical regularities in high-dimensional data, has no purchase on it. There is no distribution to approximate. There is no gradient to descend. The hardness is asymptotic; it does not bend because GPUs got cheaper or because a model memorized a billion transactions. You cannot train a model to solve ECDLP any more than you can train one to factor a prime. Context first, because the distinction matters more than the headline. ECDSA — Elliptic Curve Digital Signature Algorithm — is the industrial standard for transaction signing across the major chains. Bitcoin and Ethereum both use the secp256k1 curve. Its security guarantee is narrow and exact: given a public key, you cannot derive the private key, because doing so requires solving ECDLP, and ECDLP is computationally infeasible with classical hardware. That assumption is the load-bearing wall of the entire industry. Everything above it — wallets, exchanges, custody, DeFi, Layer 2 rollups — sits on this single mathematical bet. So when a researcher of Drake's standing invokes "bunker mode," the market should ask a forensic question, not an emotional one. Which assumption is actually under strain? What is the mechanism? Where is the proof of concept? The answer, in the available material, is: none of the above. There is no PoC, no timeline, no reproducible attack. There is a warning. That is a signal, not an alarm. Now the evidence chain. The real threat to ECDSA is quantum computing, specifically Shor's algorithm, which solves discrete logarithms in polynomial time and would let an attacker derive a private key from a public key. That is the genuine cryptographic risk. Binding "AI progress" to "ECDSA weakness" is a narrative splice. AI's indirect exposure to ECDSA is limited to three things: accelerating the discovery of implementation-layer bugs such as nonce reuse and side channels; brute-forcing weak or biased parameters; and automating vulnerability hunting. Those are engineering risks. They are not an algorithm breaking. Even the strongest model trained on every signature ever produced learns nothing about the scalar that generated them. This is where the audit work becomes useful. Based on my own contract review experience — I manually audited the Zilliqa genesis contracts during the 2017 ICO boom and found an integer overflow in the sharding transaction batching logic — I can tell you that the dangerous failures are almost never the theoretical ones. They are the implementation shortcuts. A reused nonce leaks a private key regardless of how strong the curve is. AI makes that class of bug cheaper to find. It does not make ECDLP easier to solve. The second half of Drake's warning is the part people are misreading. Wallets whose public keys are already public do face a distinct risk. Here is the mechanism. In Bitcoin and Ethereum, an address is a hash of a public key. While an address has never spent, only the hash is exposed. The moment any outbound transaction is signed, the public key is broadcast and permanently written to the chain. Metadata holds the provenance the price ignored — tracing that exposure is the whole game, following the exit liquidity to its cold storage, except here the leak is the key itself, not the coins. The implication is uncomfortable and far broader than AI. Most active wallets have already exposed their public keys. The only genuinely protected addresses are the ones that have never spent — the untouched cold addresses. If a quantum adversary ever matures, the vulnerable set is not a handful of whales. It is the entire spent-address surface of the network. That is the systemic risk worth modeling, and it has nothing to do with machine learning. The systemic risk checklist here is short and brutal. One: how much of the network's value sits in spent addresses with exposed public keys? Two: what is the realistic timeline for a cryptographically relevant quantum computer? Three: does any wallet vendor have a live key-rotation path? Four: who owns the liability if migration goes wrong? None of those questions involve AI. All of them involve custody, coordination, and code. Which brings the contrarian point. Correlation is not causation, and here the correlation is manufactured. A respected researcher raises a future risk. Media compresses it into a present danger. The word "AI" carries more emotional voltage than the word "quantum," so the splice sticks. The result is a threat model that points in the wrong direction. Attribution errors in cryptography are expensive; they send capital and engineering hours toward the wrong defense. If the industry builds its bunker against AI, it will be facing the wrong way when the actual pressure arrives from quantum migration, key rotation, and post-quantum cryptography. Watch the "bunker mode" phrase itself. It resolves into one of two engineering realities: a cryptographic emergency migration toward post-quantum schemes, or an industry-level freeze-and-migrate protocol for high-risk addresses. The source material specifies neither. The code doesn't lie, but a metaphor does — and "bunker mode" is a metaphor until someone publishes a migration path. My honest read, and I say this as someone who built wash-trading detection models on five years of on-chain data, is that the dominant risk here is narrative, not cryptographic. The technical risk is real but long-dated. The narrative risk is immediate. Expect the warning to be repackaged into anti-quantum token pitches and "AI-secure" wallet marketing before any EF roadmap appears. That is the pattern: manufactured urgency ahead of delivered infrastructure, the same shape as the "liquidity fragmentation" story VCs have pushed for three cycles. The sequencer analogy holds too. Decentralized sequencing has been a slide deck for two years; the security equivalent may follow the same arc — a compelling warning, a slow delivery. What to track next week is narrow and falsifiable. Does the Ethereum Foundation publish an official position, or does this stay a personal view? Does Drake clarify whether he means AI-assisted vulnerability discovery or a quantum threat? Does any major wallet ship a key-rotation or post-quantum migration plan? Any of those three converts a signal into an event. Until then, the ledger holds the truth and the headline holds the noise. The private keys are safe from AI. The public keys were never private to begin with.