It’s not the quantum computer that will break Bitcoin’s cryptography first. It’s the AI we’re building today.
A few weeks ago, a leaked internal memo from Anthropic’s cryptography team started circulating in private Signal groups. The details were sparse — deliberately vague, as if someone wanted to plant a signal without triggering a panic. The memo described an unexpected result: a large language model, when given access to a symbolic algebra engine, had spontaneously discovered a probabilistic attack on a lattice-based signature scheme. Not a full break. But a crack. A wedge.
No one outside Anthropic has verified it. The company has not published a paper. Yet the whisper is enough. Because if true, the timeline we’ve been using to plan Bitcoin’s post-quantum future just collapsed.
The narrative has shifted.
For years, the crypto industry has operated under a comfortable assumption: Bitcoin’s ECDSA will be broken by a quantum computer sometime after 2035, maybe 2050. The threat is distant, slow, and expensive. The solution is post-quantum cryptography — PQC — standardized by NIST, tested by academics, waiting for activation. Everyone nods. Everyone relaxes.
But the AI threat vector is different. It’s not about Shor’s algorithm or Grover’s search. It’s about neural networks that learn the algebraic structure of a cryptosystem faster than any human mathematician. It’s about reinforcement learning agents that optimize attack parameters by brute-forcing over hidden lattices. It’s about language models that can read every paper on CRYSTALS-Kyber and then ask the right questions — questions that lead to a side-channel that doesn’t require a quantum chip, just a GPU cluster.
I don’t gamble. I count the lattice points.
Let’s pull back the hood. Bitcoin currently uses ECDSA over secp256k1. That’s a discrete-log system. Quantum computers with ~4000 logical qubits can break it via Shor’s algorithm. That’s the known threat. The industry’s answer is to migrate to hash-based signatures like SPHINCS+ or lattice-based schemes like Dilithium. PQC is assumed to be safe from classical computers — they rely on hard problems that even the fastest silicon can’t solve.
But here’s the blind spot: “classical computer” is a moving target when you add AI. A GPU cluster running a transformer model is not the same as a CPU running a brute-force loop. Neural networks can approximate the solution space of a lattice reduction problem. They can learn the distribution of short vectors. They can generate pseudorandom candidates that are statistically closer to the solution than random sampling. Not a proof, but a probability.
That’s what the Anthropic memo hinted at. Their model, after being trained on cryptographic literature, identified a non-trivial correlation in the noise distribution of a lattice-based key generation algorithm. The correlation itself didn’t break the system. But it reduced the entropy of the secret by 12 bits. That’s enough to make a brute-force search feasible for a nation-state actor with access to an AI cluster.
Incentive-driven causality.
Why would an AI lab even look at this? Because cryptography is the foundation of everything. If you can crack a lattice, you can forge signatures, decrypt traffic, manipulate smart contracts. The incentive for an offensive AI is enormous. And unlike quantum computing, which requires building a completely new hardware stack, AI attacks only need software. The same GPU you use to train a chatbot can be repurposed to attack a post-quantum key exchange.
In 2020, I wrote a script to arbitrage Uniswap and SushiSwap pools. The pattern recognition that a simple Python script could perform — finding price discrepancies across liquidity pools — was already faster than any human trader. Now imagine that same pattern recognition applied to finding collisions in a hash function. The dimensionality is higher, but so is the computational power of modern models.
Code doesn’t lie, but narratives do.
We are currently in a bear market. Survival matters more than gains. But survival also means reassessing the deepest assumptions. If AI can threaten PQC before quantum computers threaten ECDSA, then the entire roadmap for Bitcoin’s security is wrong. The upgrades we’re debating today — Taproot, Schnorr, even Lightning — are irrelevant if the underlying signature scheme becomes breakable by an AI that any well-funded lab can run.
The contrarian angle: AI is also the best tool to defend against these attacks. The same models that find weaknesses can be used to verify proofs and harden protocols. The real race is not AI vs. crypto; it’s offensive AI vs. defensive AI. The Bitcoin community’s conservatism is a feature, not a bug. Rapid change introduces more risk than the hypothetical AI threat. But ignoring the signal is also a risk.
What should we watch? First, whether Anthropic publishes a paper. If they do, the narrative gains credibility. Second, whether any AI lab demonstrates a practical attack on a real-world PQC implementation, even a toy one. Third, whether the Bitcoin Core developers start discussing AI-hardened signature schemes.
Arbitrage is just geometry disguised as finance. And narrative is just geometry disguised as time.
We are preparing for the wrong boogeyman. The quantum computer is a known enemy with a known timeline. The AI that can learn to break a lattice is a ghost. It might already be in the room. The question is not if it can break crypto. The question is whether we will upgrade the protocol before the ghost learns to forge a signature.
Panic is just poor risk management. But ignoring a low-probability, high-impact event is also poor risk management. The next time you read a tweet about “Bitcoin is quantum-safe”, ask yourself: “Safe against which adversary?” The answer might not be a billion-dollar quantum computer. It might be a model trained on a dataset that cost a million dollars to assemble.
