Stanford's Dan Boneh says AI could shrink a decade of cryptographic vetting to a month
Stanford's Dan Boneh says AI could cut the roughly 10 years a new cryptographic assumption waits before deployment to about a month. On keys already in use, fellow a16z crypto panelist Eddy Lazzarin said recent AI math results showed no new weakness.
Not a builder's beat, but builders have a standing stake in it.
What happened
- The a16z crypto panel was convened after OpenAI's math results led some industry voices to warn that AI could break the cryptography in everyday use.
- Justin Thaler said cryptographers believe no efficient algorithm exists for the underlying hard problems, and no AI, however capable, can find one that does not exist.
- Boneh said the hardness of these problems cannot be proved without resolving P versus NP, and none of the newly posted AI results come close to doing so.
- Lazzarin said quantum fears have already pushed teams to treat elliptic curves as vulnerable and to deploy many of the mitigations an AI break would also require.
Why it matters
- Who has to decide decision Teams already moving off elliptic curves for quantum reasons get much of the AI protection the panel discussed from the same work, though the panel's own summary says that overlap has limits.
- What it costs cost NIST already expects cryptanalysis to force algorithm replacements from time to time. Faster AI-driven analysis would mean more frequent replacements, and systems that cannot swap algorithms while running pay for each one.
- What doesn't add up contradiction All five guests work at or advise a16z crypto, so the episode answers the warnings with one firm's cryptographers and cannot show whether the wider field is divided.
Boneh's figures come from how cryptographers build trust in an assumption: researchers try to break a design, and it gains confidence by surviving sustained attack [12]. He said that in the past "we'd have to wait like 10 years before that system could be deployed" [2]. With models in the loop, he said, "Now we can just let a model run for a month, try to break it, and if the model can't break it, boy, do we have a lot of confidence that the assumption is valid" [2]. Taking his round numbers, that shortens the vetting window by a factor of about 120 [14].
How much a failed month-long attack proves depends on how good the model is at cryptanalysis. The AI results that prompted the episode contained no cryptography questions [11]. "So you might wonder, well, does that mean the models are really bad at cryptography? We don't know," Boneh said [11]. In our view, until someone measures how well models do cryptanalysis, a month without a break shows only that an assumption survived an attacker of unknown strength.
The source did not identify the industry voices who issued the warnings [3].
What to watch
- A published measure of how well models attack cryptographic assumptions, the open question Boneh raised about model skill at cryptography.
- A named cryptographer outside a16z crypto arguing that AI has weakened a specific deployed assumption.
- NIST's final crypto-agility white paper and the virtual workshop on sector-specific strategies it has said it will hold.
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Dan Boneh is a professor at Stanford University who heads the applied cryptography group and co-directs the computer security lab; he is also special research advisor at a16z crypto.
- [2]
Boneh said: "In the past, when someone would make an assumption and build a system based on it, we'd have to wait like 10 years before that system could be deployed" and "Now we can just let a model run for a month, try to break it, and if the model can't break it, boy, do we have a lot of confidence that the assumption is valid."
- [3]
The episode was recorded after OpenAI's math results sparked warnings from industry folks that AI could break the cryptography everyone relies on.
- [4]
Eddy Lazzarin said: "We didn't learn that elliptic curves are vulnerable in a way that they weren't before. We didn't learn that any specific cryptographic assumptions have less security than they did before."
- [5]
Justin Thaler said: "We generally believe there is no efficient algorithm for these problems. ... And so if there is not one, it doesn't matter how smart the AI is. It cannot find the nonexistent algorithm."
- [6]
Boneh said cryptographers cannot prove certain problems are hard because that is related to the P versus NP problem, and "None of the problems that were posted this week come even close to resolving the P versus NP problem."
- [7]
Lazzarin said that because of quantum fears people have begun to treat elliptic curve cryptography as possibly vulnerable soon, and "we have already actually engaged in and prepared and deployed many of the mitigations you'd need for those exact concerns for a different threat."
- [8]
NIST published draft white paper CSWP 39 on crypto agility on March 5, 2025, with comments due April 30, 2025; it defines crypto agility as the capabilities needed to replace and adapt cryptographic algorithms without interrupting the flow of a running system, and NIST invites discussion of sector-specific strategies at a future virtual workshop.
- [9]
NIST says advances in computing capabilities, cryptographic research, and cryptanalytic techniques periodically create the need to replace algorithms that no longer provide adequate security.
- [10]
The other guests were Eddy Lazzarin, general partner at a16z crypto; Justin Thaler, research partner at a16z crypto and associate professor at Georgetown; Liz Harkavy, investment partner at a16z crypto; and Sonal Chokshi, editor in chief at a16z crypto.
- [11]
Boneh said of the recent AI math results: "there are no questions here on cryptography. ... So you might wonder, well, does that mean the models are really bad at cryptography? We don't know."
- [12]
Trying to break cryptography is part of how researchers build confidence in it; a design that survives sustained, serious attacks earns trust.
- [13]
Some systems at issue in the AI concerns are already subject to migration efforts because of quantum computing, and that overlap has limits.
- [14]
Boneh's round figures imply a vetting window roughly 120 times shorter.
Sources & coverage · 2 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.
- open.substack.com6d agoCan AI really break cryptography? Experts weigh in