If AI Can't Break Cryptographic Assumptions, Confidence Runs Deeper Than a Decade of Verification
A ‘bunker-mode’ warning tweet sparked panic, but experts point out: there's no evidence that cryptographic assumptions have been weakened—the real problem is that AI labs have never disclosed whether they've tried using models to attack these assumptions.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
700 papers on AI math breakthroughs exist, none address cryptography
Dan Boneh reviewed the approximately 700 papers OpenAI published and found that while they demonstrate impressive advances in mathematics and computer science, not a single one actually addresses cryptography—not even open problems in theoretical cryptography for constructing new cryptographic systems. This absence is significant precisely because it's absent. It could suggest that AI models are genuinely weak at cryptography, or it could mean labs discovered something and chose not to publish for various reasons.
— Dan BonehCryptographic confidence stems from failed attacks, not mathematical proof
Modern cryptography's foundational assumptions—like the discrete logarithm problem on elliptic curves or the shortest vector problem on lattices—have never been mathematically proven to be truly hard. Proving them hard would be equivalent to solving the famous P versus NP problem. Instead, confidence comes from decades of attempts to break these assumptions that have consistently failed. If AI labs let their strongest models attack these assumptions without success, that's not bad news—it's unprecedented validation. Even a 10-bit security reduction would be extremely valuable information. The real source of confidence has always been ‘can't break it,’ never ‘proved it's unbreakable.’
— Dan BonehAI can compress ten years of cryptographic validation into a month
Historically, new cryptographic systems required roughly ten years of community scrutiny and attack attempts before being trusted, depending on whether the global cryptographer community correctly allocated attention to important systems. Labs can now let their strongest, least-constrained models attack a candidate system for a month, and if they can't break it, the confidence gained might exceed what traditional ten-year processes accumulated. The problem isn't that AI makes cryptography riskier. The problem is that labs possess these test results internally and don't publish them. Information asymmetry is breeding the panic.
The actual risk isn't human weakness—it's whether hard algorithms exist
The discussion clarified a crucial distinction: cryptographers believe that problems like discrete logarithm don't have efficient algorithms not because ‘humans haven't been clever enough to find one,’ but because they believe no such algorithm exists. If an algorithm genuinely doesn't exist for a problem, no amount of AI capability will invent an algorithm from nothing. The real risk exposure is ‘our belief in non-existence might be wrong,’ not ‘AI is so smart it'll eventually find it.’ That's a fundamentally different kind of threat.
— Dan BonehAI poses a wider cryptographic threat than quantum computing does
Quantum computers threaten only one class of system—elliptic curve cryptography like ECDSA—which is exactly why the industry began preparing to migrate to post-quantum cryptography. But AI math breakthroughs in principle threaten a wider surface. They could equally degrade or break the lattice-based schemes designed to resist quantum computers. Since lattice cryptography is the main post-quantum candidate the industry is migrating toward, the urgent question becomes: can the latest models attack these post-quantum candidates themselves?
— Dan BonehOne lattice-cipher error destroys the system; hash signatures tolerate component failure
Blockchain communities debate whether to migrate to lattice signatures or hash-based signatures. Hash signatures offer a structural advantage: if one specific hash function is found to be weak, you swap in a different hash function and keep the same framework. Lattice schemes depend on a single shared class of assumptions, so if that foundational assumption breaks, the entire mechanism collapses with no repair path. The tradeoff is that hash signatures sacrifice lattice cryptography's richer algebraic properties—like threshold signatures and non-interactive key derivation—so cryptographers see this potential forced migration with some disappointment.
— Dan BonehAI enables near-zero-bug software, not an expanded attack surface
One optimistic counter-argument: if AI can attack cryptographic systems, startups can equally use it to scan and fix vulnerabilities in their own code before deployment. Historically, cryptographic systems often received their most damaging attacks only after deployment, the worst possible time for a breach. As the per-unit cost of intelligence dropped 90–95% over the past year, while the bug count in software is finite, we can theoretically approach ‘zero-defect’ software. This shift is more likely to raise global security than lower it.
Even if elliptic curves fail, the internet has fallback protections ready
Many mainstream internet connections, including upgraded systems like iMessage, already use ‘hybrid encryption’—encrypting the same data simultaneously with both elliptic curves and lattice cryptography. Breaking one doesn't matter because the other remains secret. If lattice cryptography also falls, the next step is standardizing a third approach. That process would be slower and messier, but the path is clear. The real risk isn't a lack of escape routes. It's whether labs will responsibly disclose any breach, giving institutions time to upgrade.
— Dan BonehIn their own words · checked verbatim
everyone calm down. Don't panic. Calm down. The world today is the same as it was a year ago, two years ago.
Dan Boneh1:01
The fact that people try to break them and fail gives us confidence.
Dan Boneh7:04
We took our best, most fully-unconstrained, most cyber- capable model, we told it to go full throttle, and we spent $20 or $50 million trying to break this scheme.
We believe that it simply does not exist.
Dan Boneh16:11
I think we humans tend to struggle with unlikely risks that seem very scary to us.
If for some reason it turns out that the basic problems on the lattices are not as difficult as we thought, the entire mechanism will collapse
Dan Boneh25:17
Then we asymptotically approach zero errors.
Figures
| OpenAI published math papers | approximately 700 | 2:04 |
| Reference scale for security degradation | 10 bits | 8:04 |
| Annual cost decline in per-unit intelligence | 90%–95% (some estimate 99%) | 35:38 |
Glossary
- Lattice-based cryptography
- Public-key cryptography built on hard problems in lattice mathematics, like the shortest vector problem; a leading post-quantum cryptography candidate.
- Cryptographic agility
- System design that allows quick switching from one cryptographic scheme to another in response to an assumption being compromised.
- SNARK recursion
- Layering zero-knowledge proofs where one proof verifies that another proof has validated a result; security properties remain incompletely understood.
- P versus NP problem
- A famous open problem in computer science whose resolution would determine whether many cryptographic problems are truly hard.
How to listen
Engineers and investors tracking blockchain cryptography migration decisions; anyone seeking to untangle AI math breakthroughs from cryptography-panic messaging.
The host's introduction of guest credentials in the first two minutes is skippable and doesn't affect understanding the subsequent argument.