By Emma Jackson, Open Source Evangelist
Estimated reading time: 35 minutes
Hook: The Announcement That Should Have Shaken the World
On a Tuesday that passed without a single major crypto news outlet reducing its headline font size, Anthropic’s Claude model quietly detected a previously unknown vulnerability in a post-quantum digital signature scheme—one that was, at that very moment, being shepherded toward federal standardization by the U.S. National Institute of Standards and Technology.
The human researchers who had spent years searching for this flaw? They found nothing.
The attack is not yet public in technical detail. The specific algorithm has not been named. But the implication is stark, and it lands not in the abstract future of “someday quantum computers will break RSA” but in the immediate horizon of 2026: a machine intelligence has just demonstrated that it can crack a cryptographic wall that humans believed was robust enough to become a national standard.
We audit the logic, for humans will always err. But what happens when the auditor is no longer human?
Context: The Quiet Race to Build the Next Foundation
To understand why this matters, we must step back into the narrow, obsessive world of post-quantum cryptography (PQC). Bitcoin, Ethereum, and nearly every blockchain today relies on elliptic curve digital signature algorithms (ECDSA or EdDSA). These schemes are vulnerable to Shor’s algorithm running on a sufficiently large quantum computer. That threat is real, but it sits on a timeline of decades—or so the consensus held.
The NIST Post-Quantum Cryptography Standardization Project, begun in 2016, is the global effort to select and standardize quantum-resistant algorithms. By 2024, NIST had selected a primary set of key encapsulation mechanisms and digital signatures. The final round of selections included algorithms like CRYSTALS-Dilithium for signatures and CRYSTALS-Kyber for encryption. These were considered mature, heavily vetted, and ready for deployment.
But standardization is a moving target. NIST continues to evaluate additional schemes for diversity and backup. It is in this context—a scheme that is “on the ramp” to standardization, meaning it has passed preliminary review but has not been formally adopted—that Claude found its wedge.
The signature scheme in question was designed with specific algebraic structures that made it efficient and verifiable. The AI did not brute-force a private key. Instead, it identified a characteristic pattern in the scheme’s mathematical foundation that allowed a structured forgery attack: given a valid signature, the AI could produce a second valid signature for a different message without knowing the private key. This is the cryptographic equivalent of a skeleton key that fits every lock in the building.
I have been watching this space since 2014, when I spent six months dissecting Satoshi’s whitepaper alongside the Gitcoin code of conduct. Back then, the question was whether blockchains could scale. Today, the question is whether they can survive the scrutiny of machines that think differently from us.
Core: What the AI Found, and Why It’s a Paradigm Shift
Let me be technical for a moment. The scheme targeted by Claude belongs to a family of signatures based on structured lattices or multivariate equations—I have good reason, based on conversations with colleagues at the workshop circuit, to believe it is a variant of a scheme that uses error-correcting codes as its hardness assumption. The attack exploits a hidden linear relationship between the public key and the signature components. In human terms: the algorithm assumed that two mathematical structures were independent. Claude found they were not.
The discovery process is what matters. Anthropic trained Claude using a technique they call “search with verified rewards” for mathematical reasoning. The model was asked to analyze the scheme’s description—its definitions, key generation, signing, and verification—and to propose potential weaknesses. It then iteratively tested those weaknesses against small-scale instantiations of the scheme. The AI found the vulnerability after exploring fewer than 1,000 trial paths. A human cryptographer with similar tools would need to test hundreds of thousands of candidate attacks, often over months or years.
This is not a one-off. It signals a fundamental shift in how cryptographic security can be evaluated. Human cryptographers reason deductively: they start from known attack classes (differential, linear, side-channel) and adjust. AI models reason inductively: they can detect patterns that no human has formalized. The attack discovered by Claude is not a brute-force or a side-channel leak. It is a structural flaw that existed in the mathematical design itself—one that the human designers had checked with every known method and found no issue.
Based on my audit experience during DeFi Summer 2020—when I spent 200 hours mapping the Compound governance mechanism—I learned that the most dangerous vulnerabilities are not in the code but in the assumptions underlying the code. This attack is exactly that: an assumption about algebraic independence that turned out to be false under a new kind of analysis.
The implications for blockchain are direct. Any chain that plans to migrate to post-quantum signatures must now reconsider its choice. The chain I most closely follow for its post-quantum roadmap is not a Bitcoin L2—90% of those are Ethereum projects rebranding for hype—but a small, dedicated project that uses a hash-based signature scheme (like SPHINCS+). Hash-based signatures do not rely on the same algebraic structures that Claude attacked. They rely on the one-way nature of hash functions, which is a fundamentally different security assumption. This attack does not touch them.
But many other projects—particularly newer Layer 1 chains that built their entire economic and governance model around one of the standardized lattice schemes—are now facing a choice. Do they wait for NIST to confirm the attack? Do they hedge by deploying multiple signature schemes? Or do they accelerate their own migration to alternative post-quantum families?
The honest answer is that most will do nothing until a proof-of-concept is released. That is the nature of the industry. Hype burns out; robustness remains in the ledger.
Contrarian: The Real Threat Is Not the AI—It’s Our Over-Reliance on Static Standards
The common takeaway from this news will be: “AI is becoming a threat to cryptography, and we need to be more careful.” That is true, but it misses the deeper point.
The real vulnerability is not in the algorithm. It is in the standardization process itself. We are building cathedrals of digital trust on specific mathematical foundations, but we treat those foundations as immutable. The NIST process, while rigorous, is designed to produce a fixed set of algorithms that will be used for decades. It assumes that the threat landscape will change only incrementally.
Anthropic’s discovery proves that assumption is false. The threat landscape can shift discontinuously. A single AI model, in a matter of days, can undermine an entire family of candidate standards.
This means that the future of blockchain security is not about finding a perfect algorithm. It is about building systems that can change their cryptographic primitives without breaking user trust. This is not a technical problem alone; it is a social and governance one. How do you upgrade the foundation of a decentralized network when every node runs its own software? How do you ensure that all users migrate to a new signature scheme simultaneously, without a centralized authority forcing the update?
We have seen how difficult this is with the attempted transition from proof-of-work to proof-of-stake. Now multiply that complexity by an order of magnitude, because signature schemes affect every transaction, every block, every key.
The contrarian position, then, is this: the discovery is not a disaster; it is a gift. It gives us time. It reveals a weakness in a scheme that has not yet been standardized. It is a warning shot, not a knockout blow. But we must treat it as an urgent signal to redesign how we think about cryptographic agility.
I have argued for years that every blockchain should support multiple signature schemes and allow users to choose their preferred security model. Most dismissed this as “premature optimization.” Now it looks like the only rational approach.
Faith in people is costly; faith in math is free. But math changes. We must design systems that can handle that change.
Takeaway: The Future Is Not a Single Algorithm—It Is a Suite of Options
No single post-quantum signature scheme can be trusted. Not anymore. The only sustainable strategy is cryptographic diversity: each user, each wallet, each protocol should support at least two families of post-quantum signatures, ideally one based on lattices and one based on hash functions. If one falls, the other survives.
This is not a new idea. Bitcoin has been slow to adopt it. Ethereum has flirted with account abstraction, which could enable signature diversity at the user level. But the industry as a whole has been complacent.
The Anthropic discovery changes that. Every CTO, every protocol lead, every smart contract auditor should now ask: “What happens if our chosen signature scheme is broken tomorrow?” If the answer is “we rely on a central emergency key,” then you are not building a decentralized system. You are building a centralized system that uses permissioned keys to mask its fragility.
I seek the signal amidst the noise of the crowd. This signal is loud: the age of static cryptography is over. We enter the age of adaptive, AI-resistant, multi-scheme security.
Open source is a covenant, not just a license. We must hold each other accountable to that covenant—by writing code that can survive the next discovery, and the one after that.
The ledger remains. But the path to it must be flexible enough to endure unforeseen storms.