Beneath the surface of the latest bull-market narrative, a quieter but more dangerous clock is ticking. In the past week, a commentary circulated through the crypto security niche citing "Anthropic's Encryption Discovery" as evidence that artificial intelligence may threaten post-quantum cryptography — perhaps before quantum computers ever dismantle Bitcoin's ECDSA. No official report, no arXiv preprint, no reproducible code. Yet the claim deserves our attention, not because it is verified, but because it exposes a fundamental blind spot in how we assess cryptographic risk. We assume the quantum threat is decades away. We assume post-quantum algorithms like Kyber or Dilithium will arrive as comfortable upgrades. But the true variable may be AI's capacity to find patterns where we see only noise, to locate flaws in mathematical structures we believed were solid. Truth is not what is seen, but what is trusted. And the market is placing its trust in a comforting timeline that may not exist.
As a decentralized protocol PM who has spent the last decade translating cryptographic guarantees into product decisions, I have learned to read threat narratives the way a nurse reads a fever chart: the temperature matters less than the trend. The trend here is twofold. First, AI capabilities are doubling in ways that directly intersect with cryptanalysis — neural networks can now reverse-engineer black-box algorithms, mount side-channel attacks from electromagnetic traces, and discover novel mathematical conjectures. Second, post-quantum cryptography, the very technology meant to shield us from quantum computers, is being treated as a solved problem. In this article, I want to put aside the hype and examine the specific mechanisms by which AI could accelerate the collapse of our cryptographic assumptions. And, more importantly, I want to ask what we should do about it — before the market's FOMO turns into a stampede of panic.
The first thing to understand is precisely what post-quantum cryptography is and why it is considered the shield. Bitcoin currently relies on ECDSA, an elliptic curve signature scheme whose security rests on the hardness of the discrete logarithm problem. A sufficiently capable quantum computer running Shor's algorithm could, in theory, derive private keys from public ones. That threat is real, but it is distant: we need thousands of logical qubits, fault tolerance, and a decade of engineering. Post-quantum cryptography is designed to remove this existential risk. The National Institute of Standards and Technology (NIST) has standardized lattice-based schemes — CRYSTALS-Kyber for encryption, and CRYSTALS-Dilithium for digital signatures. These rely on the hardness of structured lattice problems like Module-LWE and Module-SIS. The cryptography community has spent twenty years building confidence in these assumptions. That confidence is the foundation upon which the next decade of blockchain security is supposed to be built.
But here is the uncomfortable truth: the security of any cryptographic system is not just the security of its underlying math. It is also the security of its implementation, its random number generator, its side-channel leakage, and its interaction with other protocols. A quantum computer attacks the math. An AI attacks the system. AI can inspect millions of lines of code and find a subtle timing inconsistency. It can observe the power signature of a signing device and reconstruct the secret key. It can generate adversarial examples that cause a steganographic channel in a consensus protocol. None of these require solving the discrete log or lattice problems. They require an intelligence capable of exploring a far larger attack surface than any human cryptanalyst ever could.
The recent commentary's reference to "Anthropic's Encryption Discovery" is frustratingly vague. There is no publicly verifiable evidence that Anthropic has found a structural weakness in a post-quantum algorithm. But the absence of evidence is not evidence of absence. In my experience leading a ZK-SNARK integration in Berlin in 2018, our team spent three months reviewing elliptic curve implementations. We eventually found a subtle bug in the way the scalar multiplication dealt with edge-case inputs — a bug that would have allowed a deterministic nonce to leak. The review was rigorous, but it was slow. An AI model trained on cryptographic code audit findings could have detected that leak in hours, not months. And the same model, if turned toward offense, could search for similar leaks across thousands of repositories, identifying targets with the same precision a defensive auditor uses to protect them.
The Fourth Wave: AI as Cryptanalyst
To understand why AI may be the real threat to PQC, we need to look at how cryptanalysis has already been changed by machine learning. In 2019, a paper showed that deep learning could break reduced-round versions of the Speck cipher. In 2021, researchers applied transformers to side-channel traces and dramatically improved key recovery rates. Today, AI-powered fuzzing tools like Google's OSS-Fuzz with neural guidance are finding memory corruption bugs that human auditors missed. The pattern is clear: AI does not need to understand the abstract math of a cipher; it needs to learn the statistical distribution of its weakness. For lattice-based schemes, the core assumptions rely on the hardness of distinguishing a random error distribution from a structured one. But AI is precisely a distribution-learning machine. If an AI model can learn to distinguish the noise generated by Kyber from true uniform randomness — even with a slight advantage — then the algorithm's security margin shrinks. The advantage does not need to be large. It just needs to be enough to reduce the effective key size below the security floor.
There is also a more subtle pathway: AI-assisted algebraic discovery. In 2022, a well-known reinforcement learning system discovered new matrix multiplication algorithms, reducing the number of multiplications required for basic operations. The same kind of search could, in principle, be applied to lattice reduction algorithms. We already know that the best lattice reduction algorithms, like BKZ, rely on heuristics. AI could explore those heuristics in ways human mathematicians have not. The result might be a significant speedup in solving Module-LWE instances of specific parameters. The scary part is that we would not even need to fully break the math. A modest speedup in lattice reduction algorithms could reduce the security level of Dilithium-2 from 128 bits to perhaps 110 bits. For an attacker with massive GPU clusters, that changes the economics of an attack. And, as AI continues to improve, the speedup could compound.
The lattice blind spot is not just theoretical. In 2024, researchers published new techniques that improved the efficiency of lattice basis reduction using machine-learned predictors. The models were trained to predict which basis vectors to swap, cutting the number of operations needed by nearly a third. While this was demonstrated on reduced dimensions, the extrapolation is concerning. Every optimization that makes lattice attacks easier lowers the effective security of every PQC deployment. The cryptography community has historically relied on worst-case to average-case reductions to argue that random instances are hard. But the machine-learning philosophy is fundamentally different: it does not care about worst cases. It cares about the actual distribution of instances found in the wild. If the marginal distribution of parameters in Dilithium keys is slightly non-uniform, or if the error sampling entropy is imperfect, an AI will find that bias long before a mathematician would.
The Bitcoin Upgrade Dilemma
The crypto industry's response to quantum threats has always centered on one phrase: "we'll upgrade when necessary." Bitcoin upgraded to Taproot in 2021, which introduced Schnorr signatures — a step forward, but not a quantum shield. Migrating to post-quantum signatures would require a hard fork, changes to transaction formats, block size increases, and a consensus decision across miners, node operators, wallet providers, and the wider ecosystem. In the current market cycle — with ETF euphoria and institutional inflows — there is zero urgency. But if AI compresses the threat timeline from "quantum far-off" to "AI may already be working on it," then the upgrade discussion becomes an emergency.
Let's be precise about the obstacles. Post-quantum signatures are large. Dilithium signatures are around 2,400 bytes, compared to Schnorr's 64 bytes. That is a 37× increase in signature size, translating to larger blocks and higher transaction costs. Replacing Bitcoin's signature scheme would require deep changes to the UTXO model, transaction ID calculation, and validation rules. There are proposals for hybrid schemes that combine ECDSA with Dilithium, but they are not yet formalized. Meanwhile, the entire ecosystem — hardware wallets, light clients, sidechains, and Layer 2 protocols — would need to coordinate a migration. This is not a simple code change; it is a constitutional amendment for a network that values immutability above all else.
I remember auditing 12 failed smart contracts after the 2022 DeFi collapse. The common thread was not the code bugs; it was the over-leveraged architecture that ignored utility. The same mistake could happen at the protocol level: we will be so distracted by speculative yield that we will ignore the need for cryptographic agility. A rushed upgrade, without proper testing and deployment, could be more dangerous than the threat it addresses. The industry needs an upgrade path that is continuous, not panic-driven. That means we need to embed post-quantum readiness into every new protocol design today, not after the first break.
From Privacy to PQC: What My Audits Taught Me
The Berlin project is a useful lens. We integrated ZK-SNARKs to achieve sub-second privacy-preserving transaction verification. The technical challenge was not just modifying the consensus layer — it was ensuring that our privacy guarantee had no hidden holes. We reduced gas costs by 40% while maintaining ZK proofs. But the real lesson was about trust. We trusted the library's implementation of the elliptic curve, but we audited it ourselves. That audit took months. What if an AI model could have done it in a day? Defense would become vastly more nimble, but offense would become equally nimble.
The 2022 DeFi collapse taught me a different lesson. I had advocated for lending protocols that promised incredible yields. When they imploded, I retreated to a cabin in Jutland for six months. I audited those 12 failed contracts, and every single one had the same underlying flaw: the designers confused complexity for robustness. They had not tested for adversarial behavior under extreme market stress. They did not model what happens when a liquidation engine is slower than a sudden gap in price. An AI stress-testing tool could have simulated those conditions and flagged the flaw before launch. But it would also have been available to attackers, who could use it to find the most efficient way to drain liquidity. This is the dual-use future. We cannot disconnect the defense from the offense.
The Counterargument: Is This Just FUD?
But here I must resist my own narrative. The more persuasive counterargument is that the "AI breaks PQC" story is itself a new form of FUD, capitalizing on both AI anxiety and quantum anxiety to create a double layer of fear. The "Anthropic's Encryption Discovery" may be nothing more than a researcher's internal exploration, later leaked without context. If Anthropic fails to confirm it, the story will fade, and the market will forget. In the meantime, the real threat might not be AI at all — it is the lack of institutional memory for cryptographic risk. The same way the crypto ecosystem forgot about the smart contract disasters of 2020 until 2022's collapse, it will forget about PQC until the first exploit. We are not prepared not because the threat is improbable, but because our governance is too slow.
Moreover, AI could be the greatest asset we have for verifying PQC implementations. Formal verification tools, powered by machine learning, can prove the absence of specific classes of bugs. Automated theorem provers can check the correctness of cryptographic primitives. In my Copenhagen summit in 2026, we brought together regulators, developers, and civil society to draft a voluntary code of conduct for AI-crypto integration. The breakthrough was "compliance as code" — translating human-readable rules into machine-readable policies. A similar approach can define "security as code" for PQC: continuous AI-driven auditing, standardized into the development lifecycle. We should not let panic lead us to discard AI; we should institutionalize its use as a defender. The real question is whom we trust to define those standards. Trust will not come from a single lab's unverified discovery. It will come from transparent multi-stakeholder governance.
This is exactly where the blockchain industry is weakest. The Bitcoin improvement process is deliberately slow to avoid unilateral control. But that deliberation is a liability when the threat landscape changes in months rather than decades. We need emergency action plans that are pre-agreed and pre-tested. We need to know in advance which nodes will signal readiness to adopt a PQC patch, what the fallback plan is if a signature scheme is broken mid-upgrade, and how to protect users whose funds are in old addresses. These are operational questions, not ideological ones. They require the same multi-party deliberation that produced the Copenhagen consensus, but they need to happen before the first exploit, not after.
There is also a deeply human dimension here. In 2025, I led work on a decentralized identity protocol integrating AI-driven reputation scores. The challenge was preventing algorithmic bias from entrenching social inequalities. We created a human-in-the-loop verification process so that 15% of reputation updates required manual review by diverse community members. The lesson was that AI amplifies human intent, whether good or bad. The same is true for cryptanalysis. A tool that can find a bug in a lattice implementation can also find a backdoor in a malicious implementation. The ethical responsibility lies in the governance around the tool: who is allowed to use it, under what constraints, and who is accountable for the consequences.
The bull market wants us to believe that the network is safe because the price is high. But real value emerges from real trust. Trust is not a static property; it is a continuous set of practices. The market has been content to trust that quantum computers are far off because that is a comfortable story. AI challenges that comfort. It forces us to face a future where the enemy is not a machine in a lab that we can ignore, but a software system that is already all around us. The question is not whether AI will attack PQC. It is whether we will build the institutional machinery to respond at AI speed.
In practical terms, I propose five concrete actions. First, every new blockchain project should include a post-quantum migration path in its initial architecture. Second, major protocols should fund independent AI-assisted auditors to run continuous adversarial simulations on their signature schemes. Third, we need a global collaboration between AI labs and cryptography researchers to voluntarily disclose potential attack findings, similar to responsible vulnerability disclosure programs. Fourth, the Bitcoin community should start a working group on "cryptographic agility," defining parameters for a future PQC upgrade before an emergency forces it. Fifth, we should treat AI itself as the first line of defense, not just the threat — by integrating formal verification into the consensus layer. These actions are not optional extras. They are a fiduciary duty to users who place their wealth in code that we administer.
We are coding the next constitution. The transition to post-quantum security is not a technical option; it is a governance obligation. If AI discovers a weakness first, the window for response will be measured in weeks, not decades. We need to act now by funding cryptographic research, mandating formal verification, and creating an emergency consensus protocol that can upgrade Bitcoin and other blockchains without sacrificing decentralization. Truth is not what is seen, but what is trusted. The market trusts that the quantum clock is slow. Perhaps it is. But the AI clock is already running. The only way to build durable trust is to have a plan, tested and rehearsed, before the alarm sounds.