a16z crypto says the latest panic over AI cracking the cryptography behind modern digital systems is getting ahead of the actual evidence. So the firm pulled together Stanford cryptographer Dan Boneh and a16z crypto executives and researchers to sort out what is known, and what is still guesswork, after OpenAI's math results set off a new round of warnings across the industry.
The conversation included Dan Boneh, a Stanford professor who heads the Applied Cryptography Group and co-heads the Computer Security Lab, as well as a16z crypto general partner Eddy Lazzarin, research partner and Georgetown University computer science associate professor Justin Thaler, investment partner Liz Harkavy, and editor Sonal Chokshi.
Recent AI progress did not address core cryptographic questions
Boneh said the new results are a big advance in mathematics and computer science, but not in cryptography itself. At 03:00, he said: "This is an amazing advance in mathematics and computer science, but there is nothing here about cryptography. ... So you might ask whether that means these models are actually really bad at cryptography. We don't know." His point was simple: broad progress in math is not the same thing as a direct breakthrough in cryptanalysis.
Lazzarin pushed the same argument at 15:19, saying: "We have not learned that elliptic curves are more fragile than they were before. We have not learned that the security of any specific cryptographic assumption is weaker than it used to be." Put plainly, the panel said there is still no new evidence that elliptic-curve systems or any other named assumptions are suddenly less secure than they were.
Smarter AI does not imply an efficient algorithm exists
Thaler said at 16:12 that much of modern cryptography rests on the belief that some mathematical problems just do not have efficient solutions. His words: "We generally believe these problems do not have efficient algorithms. It is not that we think one might exist and humans just are not smart enough to find it. We think it simply does not exist. If it does not exist, then it does not matter how smart AI is. It cannot find an algorithm that does not exist."
Boneh added at 06:00 that cryptography still leans on hardness assumptions that cannot be fully proved today, because the issue runs straight into the unresolved P versus NP problem in computer science. He said: "We have to assume certain problems are hard. We cannot prove these problems are hard, because this is tied to a famous open problem in computer science, namely the P versus NP problem. In a sense, we would first have to solve that problem before we could really argue that certain cryptographic problems are truly, truly hard. ... Nothing announced this week came anywhere close to solving the P versus NP problem. We still have to keep making assumptions in cryptography, right?"
AI could compress years of cryptographic review
The panel said the nearer-term effect of AI may be speed. Security review, not the collapse of existing cryptographic systems. Boneh said at 12:10 that when researchers propose an assumption and build a system on top of it, the field has often waited about 10 years before deploying that system so other people have enough time to analyze the assumption and build confidence in it.
He said: "In the past, when someone proposed an assumption and built a system based on it, we would often wait about 10 years before deploying that system, in order to build confidence and give people enough time to analyze the assumption. ... But now, all of a sudden, we can use these models to speed up that process dramatically. ... We can now let the models run for a month trying to break it, and if the models cannot break it, then our confidence that the assumption is valid rises substantially."
That puts AI in a very different box. Not proof that cryptography has broken. More like a tool that could cut down the time needed to attack, test, and review new designs.
Quantum-readiness work may help with part of the AI risk
The speakers also said some of the prep work for AI-related worries is already in place, because the industry has spent years getting ready for quantum threats. Lazzarin said at 20:52: "Because of concerns about quantum threats, people have already started treating elliptic-curve cryptography as something that could soon become vulnerable. So for these exact same concerns, we have actually already been involved in, prepared, and deployed many of the mitigations you would need, just for a different threat originally."
But Boneh said at 22:08 that the overlap only goes so far. He said: "In a sense, the AI threat is broader, because it could also examine our post-quantum systems and somehow weaken them, or even break them entirely."
Cryptographic agility remains a key defense
Boneh said at 22:47 that cryptographic agility gives deployed systems a way to switch from one design to another if the evidence changes or a method is shown to be insecure. He said: "The systems we deploy can quickly switch from one design to another when the design they are using is shown to be insecure. In a sense, cryptographic agility is our friend in defending against surprises."
He added at 38:40 that internet connections are already using mixes of cryptographic methods: "They are actually already using agile cryptography. They use both elliptic curves and lattices for encryption at the same time. If you break elliptic curves, that is fine, because it is also encrypted under a lattice-based system."
The clearest near-term use case is finding flaws before release
For builders, the panel said one of the most obvious openings is bringing more AI-assisted security testing into development before software goes live. Talking about complex cryptographic systems, Thaler said at 28:56: "The attack surface for vulnerabilities becomes very large, yes, and I am very optimistic that these AI tools that are making people panic about cryptography will also help us eliminate all the vulnerabilities. This is a very hard problem."
Boneh said at 35:05 that once a system is deployed, it naturally draws attention from many cryptographers, but by then attacking the system may already come too late. He said: "After a system is deployed, it attracts the attention of a large number of cryptographers, but by then attacking the cryptographic system is actually too late. And now, because we control these AI models, any startup deploying new cryptographic capabilities, or even any company deploying any software, can have these models analyze vulnerabilities in its cryptography and in its software as a whole."
Lazzarin added at 37:10: "Vulnerabilities can be eliminated. They can be avoided through design, they can be mitigated, and the main resource we use to mitigate them, AI capability, is getting dramatically cheaper every few months."
The original article said the quoted material had been lightly edited for clarity, and that ellipses mark omitted passages.