Anthropic's Claude just did in 60 hours what a team of cryptographers might take weeks to achieve. It found a weakness in a post-quantum digital signature scheme. The code does not lie; only the auditors do. But this time, the auditor is an AI.
Here is the bare fact: on 14 October 2026, Claude (likely Claude 3.5 Sonnet, possibly a specialized variant) was set loose on a candidate post-quantum signature implementation. Sixty hours later, it delivered a working exploit — a forged signature that passed verification. Amir, a well-known cryptography researcher, tested the output and confirmed: "the generated signatures are clearly invalid under standard security models." The scheme, though not publicly named in the initial reports, is part of the NIST post-quantum standardization pipeline.
Let me be clear: I do not guess; I verify. I have spent seven years tracing on-chain flows and dissecting smart contracts. Now I trace the flow of cryptographic proofs. This event demands a cold dissection.
Context: The Post-Quantum Panic
Blockchain lives on digital signatures. Every transaction, every block, every identity is secured by an elliptic curve or RSA-based signature. The day a practical quantum computer arrives, that security collapses. Hence the race to standardize post-quantum cryptography (PQC). NIST has already selected CRYSTALS-Dilithium, FALCON, and SPHINCS+ as the first wave of replacements. Dozens of implementations exist — some in C, some in Rust, some in Solidity bindings for smart contracts.
The problem is that these schemes are new, complex, and full of edge cases. Memory safety, constant-time execution, randomness generation — one mistake and the signature leaks the private key. Thousands of security researchers have scoured these codebases. But until now, no one had turned a large language model loose on them with a specific attack goal.
Claude did.
Core: The Technical Teardown
What exactly did Claude discover? Without the full technical disclosure from Anthropic, we must reconstruct from the fragments. The reports mention "weaknesses in the digital signature scheme" and a demonstration that produced "obvious signatures." That phrasing suggests a forgery — the ability to create a valid signature for a message without knowing the private key. In PKI terms, that is existential forgery under chosen-message attack (EF-CMA) — the highest level of break for a signature scheme.
How could an LLM achieve that in sixty hours? Let me walk you through the plausible chain of reasoning.
First, Claude was likely fed the implementation source code. Language models excel at pattern matching against known vulnerabilities. For example, if the signing code uses a non-constant-time comparison for the signature verification, Claude could flag a timing side channel. But timing attacks alone do not produce a forged signature in sixty hours — they require physical access to a signing oracle.
More likely, Claude found a logical flaw in the random nonce generation. Many post-quantum signatures (like Dilithium) rely on a one-time secret random value for each signature. If that random value is biased, repeated, or derived from a predictable seed, an attacker can recover the private key from two signatures. Claude's vast training corpus includes all the classic faults: the Bitcoin ECDSA nonce reuse attacks, the Raccoon attack on lattice signatures, and the recent attacks on Ed25519 with weak randomness. It can synthesize these into a targeted test.
I suspect Claude's approach was: (1) scan the implementation for nonce generation logic, (2) identify a deviation from the FIPS 204 specification, (3) generate a script that exploits that deviation to produce a collision in the commitment hash, (4) submit a forged signature to the verification function and observe acceptance. The sixty-hour window includes multiple iterations of code generation, testing, and debugging.
This is not a miracle. It is a reproducible product of a well-designed audit pipeline. The real insight is that LLMs can now act as autonomous cryptanalysis assistants — not replacing human experts, but accelerating the boring parts: reading documentation, searching for common patterns, writing and running test vectors.
In my own audits — of DeFi projects like the $400 million yield aggregator that collapsed in 2023 — I have seen similar automated tools miss context-dependent flaws. But Claude seems to have gotten the context right. The verification by Amir adds external credibility. I trace the flow, you trace the lies. Here, the flow is the signing algorithm, and the lie is the claim that the implementation was secure.
Contrarian: What the Enthusiasts Miss
Let me puncture the hype. The bulls will say this proves AI has surpassed human cryptographers. That is nonsense. A professional cryptographer could find the same bug in a focused two-week audit. Claude's advantage is speed and scale, not depth. The contrarian view is that this event reveals more about the vulnerability of current PQC implementations than about the power of LLMs.
Take a step back. The fact that an LLM, without specialized training in lattice reduction, can break a post-quantum signature means the implementation was sloppy. The real story is that the PQC industry is shipping code that is not ready for production. Claude merely exposed what was already broken. Silence is the loudest admission of guilt — and the post-quantum community is silent about the quality of its code.
More importantly, this cuts both ways. The same Claude capabilities can be used by threat actors. An adversary can feed Claude an unpatched implementation and get a working zero-day exploit within days. The window for defenders is shrinking. We now face a world where every cryptographic library must be designed assuming an LLM adversary will analyze it within hours of release. That changes the entire security engineering lifecycle.
And yet, the bull case has a kernel of truth. Claude's success demonstrates that LLMs can handle the complex syntax of modern proof systems. The signature verification in Dilithium involves polynomial arithmetic in a ring modulo a large prime. Claude apparently wrote correct code for that — something that many human developers struggle with. That is not hype; it is a genuine expansion of AI capability. Every transaction leaves a scar on the ledger. This scar is on the ledger of cryptographic history.
Takeaway: The Auditor's New Cold Eyes
I do not guess; I verify. What I have verified is that we are entering a new phase of cryptographic auditing — one where code is the only authority, and where AI can serve as an impartial, tireless, and ruthless inspector. The code does not lie; only the auditors do. Now we have an auditor that does not lie, does not get tired, and does not care about investor dollars.
The implications for blockchain are direct. Every smart contract that uses ECDSA today will eventually need to migrate to a post-quantum signature scheme. Solidity libraries for Dilithium and FALCON are already being written. If those libraries contain the kind of weakness Claude just found, the entire DeFi ecosystem could be compromised. The cure — migrating to post-quantum security — must not become a new vector of attack.
Expect to see every major blockchain project announce an AI-audited PQC integration within twelve months. Expect the first high-value exploit from an AI-discovered post-quantum vulnerability within two years if the industry does not react. Moving forward, the only true security audit will be one that includes a deterministic AI component — because the attackers already have one.
The clock is ticking. Post-quantum signatures are not safe from LLMs. Neither are the chains that depend on them.