Matthew Green's warning is getting attention for a specific reason. The Johns Hopkins cryptographer isn't predicting that public key infrastructure (PKI) is about to collapse. He's pointing at what it's built on. Machines have started outperforming humans on some mathematical problems, and much of today's public-key cryptography rests on a small number of hard problems that have been studied for decades. That's a narrower exposure than "encryption is broken," but it's the kind of exposure that matters.
In a series of posts on X, Green put a rough number on his concern. "I think there is a 1% chance we live in Minicrypt, and a 15% chance we functionally lose confidence in our existing public-key encryption algorithms," he wrote. Minicrypt is the cryptographers' term for a world with one-way functions but no public-key cryptography. A 1% chance of that world is a low number. A 15% chance of losing confidence in the algorithms we use now is not.
His follow-up made the target clearer. "I'm worried about standardized public-key encryption (and some signatures) and specifically the fact that most of it is based on a small number of hard problems," he added. The concern isn't that TLS or code signing will stop working tomorrow. It's that the whole stack leans on a handful of mathematical assumptions, and an assumption that looks solid for years can crack in a single result.
Why decades of study may not be enough
The usual comfort is that these problems have been picked over by human mathematicians for a long time. Green's point is that this history no longer guarantees safety, because the machines doing the picking have changed.
Markku-Juhani O. Saarinen, a cryptographer and professor at Finland's Tampere University who has studied AI-assisted cryptanalysis closely, draws a useful line between two kinds of hard problems. The first is numerical: problems where a solution is believed to be expensive for computers to find. The second is analytic: open questions that a mathematician would normally settle with a proof, a counterexample, or a new theory.
Saarinen argues the second category is where the shift happened. "What has happened in 2026 is that mathematics had its 'Lee Sedol moment' (after the 'last Go grandmaster') and [analytical] problems previously thought to be in the human domain (2) are now being solved by computers," he said.
That distinction matters for the cryptography itself. Schemes built on factoring large primes, discrete logarithms, or learning with errors (LWE) depend on numerical hardness. An analytic breakthrough doesn't have to make those problems trivial to hurt a scheme. It only has to shave the effective difficulty. As Saarinen put it, "The fear in cryptography is that AI will find a shortcut for some hard computational problems and make them easy. While AI is very good at optimizing computation, we don't expect it to turn all hard computational problems into trivial ones."
What has already broken
The evidence isn't purely theoretical. Saarinen points to HAWK, which an AI system helped break in 2026. HAWK is a relatively new scheme, but he notes that numerous older proposals have also fallen to human analysts over time. A cryptographically relevant break doesn't have to be fatal to be important. As Saarinen explained, "A cryptographically relevant break can be exponential but might halve the number of bits."
AI also played a significant role in finding a major shortcut against the McEliece cryptosystem, one of the earliest public-key schemes, dating to 1978. That one hits closer to home for standards bodies. Saarinen notes that Germany's federal cyber agency had recommended a variant, Classic McEliece, for quantum security since 2020, then withdrew that recommendation at the end of September.
His read on the damage is measured. In both cases, he says, the schemes could still be "considered secure if security parameters were adjusted, but that would have doubled public and secret keys, signatures, and ciphertexts." In other words, the fix exists, but it isn't free. Bigger keys and bigger ciphertexts are real operational costs for anyone who has to deploy them at scale.
Saarinen is also blunt about the older systems. From a computational-complexity standpoint, RSA-style and elliptic-curve cryptosystems are already effectively broken against a sufficiently large quantum computer, since "polynomial-time quantum algorithms for them have been known since 1994, and building quantum computers is 'just an engineering problem.'" That's a quantum argument, not an AI one, but it explains why the industry is already moving. The main post-quantum candidates, ML-KEM, ML-DSA, and SLH-DSA, have received more than a decade of intense analysis.
Public-key versus symmetric: where the shortcut hides
A natural follow-up question is whether AI is more likely to threaten public-key or symmetric cryptography. Saarinen's answer leans toward public-key. Public-key schemes need mathematical structure to create a trapdoor, meaning a secret that lets someone decrypt a message or sign one in a way the public key can verify. That structure is exactly what a clever analytic attack can grab onto.
"Symmetric-key cryptography (hashes and keyed ciphers) has simply more design freedom in this sense," Saarinen concluded. Symmetric primitives aren't free of risk, but they don't need to encode a hidden algebraic relationship, so there's less for an attacker to find.
Q-Day, LLMs, and the case for agility
Not everyone thinks large language models are already reshaping cryptanalysis. Alan Woodward, a computer scientist at the University of Surrey, says LLMs haven't made a huge dent yet. He does see erosion. "We've seen some inroads and erosion of the security guarantees, but what's actually happening as far as I can see is that the LLMs are acting as a great tool for mathematicians, cryptographers, and computer scientists to speed up their research," he said. "That leap of intuition needed to find the fatal flaw still seems to need humans."
His view on where the real weakness sits is pointed. "LLMs are very effective at finding flaws in those implementations rather than the underlying mathematics," Woodward said. That's a different risk from a broken theorem. Bugs in implementations are common, and they're often easier to fix than the math.
Woodward draws a parallel between AI-driven cryptanalysis and the quantum threat to legacy schemes. Both call for cryptographic agility. "We've already seen people migrating in preparation for 'Q-Day' so in some ways that should already be addressing the risks posed, if they're real, by LLMs," he said. His preferred approach is hybrid: "a mix of cryptographic schemes that, whilst one or two might fall away, leaves us with others we can rely upon. Any found to be flawed for whatever reason can be deprecated rapidly and the intact schemes used instead."
Saarinen lands in the same place, though with a sharper warning about overreaction. "We really can't entirely rule out surprises, but I'd also definitely not recommend some hasty moves, because those are sure to cause problems," he said. "Cryptography has good standards, best practices, and active modernization paths (the 'PQC transition') — just following those is the most sensible approach from information security perspective."
Green's concern is the same underlying risk seen from the other side. A cryptanalytic surprise that arrives faster than standards and deployments can adapt is the scenario that could really shake confidence in internet security. Green hasn't yet responded to requests for comment on the debate his posts sparked, and the discussion on LinkedIn and elsewhere is still running. The practical takeaway for security teams is that the question isn't whether to abandon PKI. It's whether your systems can swap algorithms, key sizes, and signature schemes quickly when a break lands. If the answer is no, that's the gap to close now.