a16z crypto Panel Discusses Whether AI Math Advances Threaten Modern Cryptography

The starting point of this discussion is not a successful break, but an absence from a list. OpenAI previously released hundreds of mathematics research manuscripts generated by internal models at once, covering multiple directions in mathematics and computer science, yet it did not list cryptography results. After reading the research list, Boneh was especially concerned about this gap—cryptography likewise has a large number of open problems, and the mathematical reasoning required is not far from some of the published results.

The gap can have different explanations. The model may have tried and never obtained a result worth reporting; the lab may also have discovered security-related content and chosen to delay disclosure. Thaler believes that in either case, researchers need to know what was tested, how many resources were invested, and how the results were handled.

Two things need to be distinguished here: the public results did not present an actual break of mainstream cryptographic schemes, which is a confirmed fact; while "whether the model tried and why there were no results" belongs to undisclosed information. Reading the former as "cryptographic systems are safe" or the latter as "it has already been broken but was concealed" both go beyond the existing evidence.

Boneh's demand therefore lands on a specific action: frontier labs should disclose whether they have ever tested deployed systems with their strongest models, and what the results were. This is not a demand to publish attack methods, but a demand to publish the existence and boundaries of the testing.

Public-key cryptographic schemes usually presuppose that certain computational problems are difficult to solve quickly. Taking the elliptic curve discrete logarithm problem as an example, a user can publish one key for others to encrypt information or verify signatures, while retaining another private key; security depends on an attacker's difficulty in deriving the secret from public information.

But for a mathematical discovery to become a real attack, it must also be turned into computational steps against specific algorithms and parameters, and be completed within an affordable cost. Even if a faster method is found, one must judge how much speed it improves. Cryptographic systems care about the actual threshold of time and computing power; there is still a distance between theoretical progress and stealing a key.

Strictly proving that an efficient solution to the relevant problem never exists involves complexity problems in computer science that remain unsolved. The cryptography community has therefore long adopted another testing method: make the scheme public, invite researchers to continuously search for faster attack algorithms, and then judge whether it is robust enough based on results accumulated over many years. Boneh said that a new scheme in the past might have to undergo nearly a decade of analysis before gradually gaining the confidence needed for deployment.

AI may compress this process. If a model can continuously search for attack paths, a research team can conduct intensive testing of a new scheme in a relatively short time; even if no break is found, the results can provide new evidence for evaluation. But its weight depends on model capability, testing scope, and invested resources, and these conditions need to be published together. To make intensive testing usable security evidence, the report must explain the tested algorithm, parameters, compute budget, and attempted attack directions.

This is precisely the crux of the information gap: model capability, testing scope, and invested resources have not yet been published together, so the weight of intensive testing results as security evidence cannot be determined. What kind of testing a scheme has undergone directly affects what institutions upgrade first and what they upgrade later; if labs publish only mathematical results, outsiders cannot use them to rank the priorities of cryptographic migration.