Add ‘post-AI’ encryption to your to-do list

Even as federal agencies move their most sensitive systems to new forms of encryption following an order from the White House in June, the release this month by OpenAI of new results to theoretical math problems has sparked concern about whether even proposed post-quantum encryption will be safe from increasingly capable AI models.

Encryption is what keeps online data secure; encrypting a message means applying mathematical equations to disguise its contents to an attacker. But new math techniques could let an attacker decipher encrypted messages without a key: the equivalent of a safecracker using the click of pins in a lock to find the right combination. If existing encryption algorithms fall, huge swaths of the internet would be at risk of attack. 

Plans for post-quantum encryption may need to incorporate stronger multi-layered defenses—and move forward more quickly than the White House anticipated—to protect data that hasn’t yet been collected.

Encryption works now because, in practice, there are ways to make it very easy for someone with the key to apply some math and reveal the original message, but not to find the key or solve the equations by brute force in a reasonable amount of time. But the encryption techniques currently in widespread use haven’t been proven theoretically safe; they’re trusted because nobody has been able to crack them so far. So if new math is discovered that makes solving encryption schemes much easier, the foundations of the internet’s security would be at risk. 

AI systems have been churning out math discoveries at an unprecedented rate. In May, an OpenAI model disproved the Erdős unit distance conjecture, a well-known problem in a subfield of math called discrete geometry. And in September, the company announced its AI had a solution to the Navier-Stokes equations. The equations make up one of seven Millennium Prize Problems, some of the most important unsolved problems in mathematics. 

“It just solved 400 problems, any of which, if a human had solved it, that could have been the highlight of their career, that could have been their tenure case,” Scott Aaronson, author of Quantum Computing Since Democritus and Chair of Computer Science at UT Austin, said in an interview.

Omer Reingold, computer scientist and co-winner of the Gödel Prize in 2009, wrote before the latest release that, “The world of mathematics, including theoretical computer science, is in turmoil. Even the Millennium Prize Problems…are no longer beyond the reach of LLMs.” 

The OpenAI papers showed new results, “in just about all areas of math, but with this very conspicuous exception that there was nothing in cryptography. And that can't possibly be an accident,” Aaronson said.

Justin Drake, a security researcher at the Ethereum Foundation, wrote that he believes there was “backroom interventionism.” 

“I've witnessed first-hand the U.S. government censoring academic quantum cryptanalysis results. Backroom interventionism is my base case,” he said.

Regardless of whether OpenAI redacted encryption results from this particular release, the flood of new results has prompted fears that AI systems could find weaknesses in existing encryption algorithms. The cryptocurrency industry, which depends on encryption to safeguard its digital currencies, has been particularly alarmed. 

As part of the way cryptocurrencies function, many crypto accounts expose some information about themselves when they send money. On X, formerly known as Twitter, Drake wrote, “Today I call upon the blockchain industry to calmly begin planning for ‘bunker mode’. My personal recommendation is to set in motion a controlled mass migration of assets to fresh addresses.” 

A “fresh address” is one that has never been used to send money to another account, limiting the ways future attackers might break its security.

Other cybersecurity experts are skeptical that AI will have much of an effect on concrete security issues. “Cryptographers are incredibly conservative... we put a lot of buffer into [secure encryption], an enormous amount,” said Bruce Schneier, security guru and author of the bestseller Applied Cryptography. 

His concern is more on implementation questions in cybersecurity than the possibility of breakthroughs in theoretical math. 

“Computer security is so bad. That's the worry. Finding vulnerabilities in code, hacking into systems. Those are the worries, not the math ... Of all the security things going on, math is the strongest we have,” Schneier added.

Emily Wenger, an assistant professor of electrical and computer engineering at Duke University, who has used machine learning to test the security of post-quantum encryption methods, said she doesn’t think “advanced large language models are going to break all encryption.”

“But when you make practical implementation choices... we found in the past that certain instantiations of these algorithms are insecure,” she said. “There's kind of two ways to break these things. You can either prove something about the math problem, or you can look at implementation choices.”

Aaronson noted that undermining the math behind widely used encryption algorithms like RSA would have sweeping consequences. “These mundane vulnerabilities, you can play the game of whack-a-mole. As soon as you find one, you patch it. Whereas if there's a break of RSA ... if all of our public key crypto systems were broken, then we would really have a problem, the way that we have constructed the internet.”

Dan Boneh, a cryptography expert and professor of computer science and electrical engineering at Stanford, said in a video that losing current forms of encryption “would affect the entire internet... it affects the entire digital world. So, home banking, healthcare, government communications, everything [would be] impacted by such a break.”

But some experts are optimistic that security will ultimately become stronger from the increased attention. 

“In the past, when someone would make an assumption and build a [cryptography] system based on it, we'd have to wait 10 years before that system could be deployed, to give people enough time to analyze it. And now all of a sudden we can accelerate this process tremendously using these models... 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. So we actually have the prospect to have more confidence in our security than ever before,” Boneh said.

Schneier concurred, saying, “It's an arms race. I think AI is the first technology that's potentially flipping that arms race to a defender advantage … Microsoft used AI to patch 900-something vulnerabilities for a Patch Tuesday last month. But they will never again be problems. You're already being attacked at computer speeds. Now you can be defending at computer speeds.” 

Still, any transition would be costly. For a typical commercial firm, building “crypto agility,” or the ability to quickly switch the organization’s data to new types of encryption, may be the best response to encryption uncertainty. But for the defense industrial base, adversaries have been collecting their online data in its encrypted form for years, so switching now would have little effect on their ability to decode existing data. 

Intelligence agencies both foreign and domestic have long embraced “harvest now, decrypt later” as a method of espionage, collecting encrypted data over time and storing it, in preparation for breakthroughs that would allow them to crack existing encryption. 

“You have to assume that intelligence agencies are vacuuming up your data now, and that's all the more true if you think that AI could potentially discover breaks for these systems within the next year or two,” Aaronson said.

For organizations that deal with restricted or classified information, the prospect of AI math breakthroughs adds another ticking clock to the post-quantum future many have only begun to prepare for.

The federal government is targeting a comprehensive move to post-quantum encryption in the coming years. Agencies must follow Executive Order 14412 and OMB memo M-26-15, requiring them to migrate key systems to post-quantum cryptography in stages from Dec 2030 to 2035. The NSA has its own schedule for national-security systems. Contractors would also be subject to NIST’s cryptographic standards by Dec 2030 under the EO. 

But new math results relevant to encryption could require a different approach. Aaronson said of AI undermining classical encryption or post-quantum algorithms, “Either of those could fall without the other. There's no logic.”

The prudent move for organizations may be “crypto agility,” the capacity to swap and layer different algorithms on top of each other as needed. With a proper agility posture, even broken encryption algorithms would not result in widespread harm in most cases. 

“You should be getting crypto agility for post-quantum. So you should be doing it already. And if you're not, you should,” Schneier said.

Matthew Green, an associate professor of computer science at the Johns Hopkins Information Security Institute and author of the blog A Few Thoughts on Cryptographic Engineering wrote, “I can’t see the future. Nobody can. But so far, betting against models and in favor of human intelligence being the limit has not worked well for any of us.”