Andreessen Horowitz’s latest crypto podcast argued that recent AI progress in mathematics has not reached the core problems that modern cryptography depends on, pushing back on concerns that new models could suddenly break the security foundations of digital systems.
The discussion came after OpenAI’s latest math results triggered fresh warnings inside the industry that AI might undermine cryptography. In response, a16z brought together Stanford cryptographer Dan Boneh and members of its crypto research team to separate what is known from what remains speculative.
The guests and the premise of the discussion
According to the a16z crypto Podcast recap, the conversation featured Dan Boneh, a Stanford University professor who leads the Applied Cryptography Group and co-leads the Computer Security Lab; Eddy Lazzarin, special research adviser at a16z crypto; Justin Thaler, research partner at a16z crypto and associate professor of computer science at Georgetown University; Liz Harkavy, investment partner at a16z crypto; and Sonal Chokshi, editor-in-chief at a16z crypto.
The starting point was a wave of anxiety tied to OpenAI’s recent math breakthrough. The speakers addressed whether AI is actually getting closer to breaking the cryptographic assumptions that support the internet and crypto systems, and which worries are grounded in evidence.
No direct hit on cryptography
Boneh said at the 03:00 mark: 「This is remarkable progress in mathematics and computer science, but there is nothing here about cryptography. ... So you might ask, does this mean these models are really bad at cryptography? We don’t know.」
Lazzarin said at 15:19: 「We have not learned that elliptic curves are more fragile than before. We have not learned that the security of any specific cryptographic assumption is worse than it used to be.」
The core point from the panel was straightforward. Recent AI results may matter for math and computer science, but they do not amount to evidence that modern cryptographic assumptions have already weakened.
If no efficient algorithm exists, AI does not change that by itself
Thaler said at 16:12 that a large share of modern cryptography rests on the belief that some mathematical problems simply do not have efficient algorithms.
His wording was: 「We broadly believe that these problems do not have efficient algorithms. It’s not that we think they might exist and humans just aren’t smart enough to find them. We think they simply do not exist. If they do not exist, then it does not matter how smart AI is. It cannot find the algorithm that does not exist.」
Boneh added at 06:00 that cryptography has to assume some problems are hard, but the field cannot formally prove that hardness because doing so would connect to the famous unsolved P versus NP problem in computer science.
He said: 「We have to assume some problems are hard. We can’t prove these problems are hard because this ties into a famous open problem in computer science, namely P versus NP. ... Nothing released this week comes anywhere close to resolving P versus NP. We still need to make assumptions in cryptography, right?」
In the panel’s framing, AI’s latest progress remains far from overturning the difficulty assumptions that support modern cryptography.
AI could compress a 10-year review cycle
The guests said the more immediate change may be speed. Rather than directly breaking cryptography, AI could sharply accelerate the process used to test and attack new assumptions and systems.
Boneh said at 12:10 that when someone proposes an assumption and builds a system on top of it, the field has historically needed around 10 years before deployment to build confidence through sustained analysis.
He said: 「In the past, when someone proposed an assumption and built a system based on it, we often had to wait around 10 years before deploying that system, in order to build confidence and give people enough time to analyze the assumption. ... Now, all of a sudden, we can use these models to accelerate that process dramatically. ... We can now let models run for a month trying to break it, and if the models can’t break it, our confidence that the assumption holds increases a lot.」
That shifts AI into the role of a testing tool, not proof that the underlying assumptions have already failed.
Quantum preparation has created a buffer, but not a full answer
The speakers also said some of the current concern around AI overlaps with work that the industry had already started because of quantum-computing risks.
Lazzarin said at 20:52: 「Because of concern over quantum threats, people had already started to view elliptic-curve cryptography as something that might become vulnerable. So for these exact same concerns, we’ve actually already been involved in, prepared, and deployed many of the mitigations you would need, except they were originally for a different threat.」
Boneh drew a limit around that comparison at 22:08. He said: 「The AI threat is broader in a sense, because it could also inspect our post-quantum systems and somehow weaken them, or even break them entirely.」
That means prior quantum-migration work may help, but it does not fully settle the AI question.
Cryptographic agility as a practical defense
Boneh repeatedly pointed to cryptographic agility, meaning a system’s ability to switch cryptographic methods when evidence changes or when a design is shown to be unsafe.
At 22:47, he said: 「The systems we deploy can switch quickly from one design to another if the design they use is shown to be insecure. Cryptographic agility is, in a sense, our friend in defending against surprises.」
He added at 38:40 that internet connections are already using this approach in practice: 「They are in fact already using agile cryptography. They use elliptic curves and lattices together for encryption. If you break elliptic curves, that’s okay, because the traffic is also encrypted under a lattice-based system.」
In that view, the key is not to assume one scheme will remain untouched forever, but to make sure systems can adapt quickly if weaknesses appear.
The clearest near-term value: finding bugs before deployment
For builders, the panel said the most concrete opportunity may be to use AI earlier in development to expand security testing and spot weaknesses in cryptographic designs and software implementations before launch.
Thaler said at 28:56, while discussing complex cryptographic systems: 「The surface area for vulnerabilities becomes very large. Yes, I’m very optimistic that these AI tools that are making people panic about cryptography will also help us clear out all the vulnerabilities. This is a very hard problem.」
Boneh said at 35:05: 「Once a system is deployed, it gets a lot of attention from cryptographers, but by then attacking the cryptosystem is really too late. Now, because we control these AI models, any startup deploying new cryptographic capabilities, and really any company deploying any software, can have these models analyze vulnerabilities in its cryptography and in its software more broadly.」
The recap also cited Boneh as saying vulnerabilities can be eliminated, avoided through design, or mitigated, though the supplied source text cuts off before the end of that quote and does not include the full wording.
Across the discussion, a16z and Boneh did not conclude that AI has already destabilized modern cryptography. Their stated position was narrower: there is no evidence in the latest results that specific cryptographic assumptions are weaker than before, while AI may prove far more useful as a tool for analysis, testing, and vulnerability discovery.