Contact Us

Y2K Without the Countdown Clock: AI, Math, and Your Encryption

Chris Clements, VP of Solutions Consulting & Client Experience

“You don’t need a math breakthrough to change the economics of an attack.”

Key Takeaways

  • AI-assisted research knocked a post-quantum candidate out of NIST’s process. These findings do not demonstrate a practical break of RSA-2048 or full-round AES. 
  • The bigger shift is economic. AI helped make known attacks cheaper to build and run. If you still have 1024-bit RSA around, you’re out of excuses. 
  • Rotating keys doesn’t change the math. Signing needs its own plan, especially for firmware. 
  • Find out how long it would take to replace a broken algorithm in your critical services. Run a fire drill and measure it now. 

Remember Y2K?

Those of us old enough to remember Y2K remember a lot of very expensive consultants and a lot of very tired COBOL programmers. We also remember that, for the most part, nothing happened. 

That wasn’t luck. People spent years digging through code and fixing it before a deadline everyone knew: January 1, 2000. You could budget for it and yell at vendors about it. 

Our current cryptography may have its own Y2K problem. Nobody gets to know the date. There may never be one, but if it arrives it might not come with a press release. 

So how long would it take your organization to stop relying on an algorithm that became unsafe tomorrow? 

The ability to make that change while keeping systems secure and running is called cryptographic agility. 

Math Got Weird This Year

If you haven’t been following AI labs’ math announcements, it’s been a strange few months. OpenAI claimed a resolution of the Navier–Stokes problem (the Clay Institute still lists it as “Active,” which I read as “settle down, we’re checking your work”). Anthropic reported raising the known lower bound for Riemann zeta zeros on the critical line from 41.6% to 67.2%. OpenAI also says it disproved the Exponential Time Hypothesis, with academic collaborators working to verify the claim. 

If you work in security, you probably had the same reaction I did. Cool… So what happens to the math that holds up all our encryption? 

HAWK Didn’t Make It

In July, Anthropic reported that Claude developed an improved attack on HAWK, a post-quantum signature scheme under evaluation by the National Institute of Standards and Technology (NIST). The research took about 60 hours and roughly $100,000 in API costs. The HAWK team withdrew their submission the next day. 

Before anyone panics: that $100K paid for research, not somebody’s production key. The attack is specific to HAWK. It does not demonstrate a break of ML-DSA, Falcon, or lattice-based cryptography as a whole. 

Still, an AI found a serious structural weakness in a scheme that professional cryptographers had been picking apart (see where I’m going with this now?). 

AES Got a Scratch

The same research improved an attack on AES-128 by 200 to 800 times. That sounds terrifying until you read the fine print: seven of AES-128’s ten rounds, and 2^105 chosen plaintexts. In other words, it isn’t a practical attack on deployed AES. Please, for the love of God, don’t put “AES is 800x weaker” in a board deck. 

RSA Records Fell, and Nobody Invented Anything

Yeah, this is the one I’d pay the most attention to. 

In September, Cognition factored RSA-260, an 862-bit number, using existing methods adapted for graphics processing units (GPUs). In fact, there was essentially no new math involved. Later that month, Stephen Weis factored RSA-896 with Claude helping adapt existing tools to GPUs. Weis was clear that this does not compromise deployed RSA-2048 keys. 

He’s right, the gap to 2048 is still enormous. AI helped make an attack everyone already knew about cheaper to build and run. 

You don’t need a math breakthrough to change the economics of an attack. Sometimes you just need the grunt work to get easier. 

Also, if you’ve still got 1024-bit RSA hanging around anywhere (DKIM keys, I’m looking at you), you just ran out of excuses. 

About That Warning

Most security programs rest on an assumption nobody ever writes down: if the math under our encryption breaks, we’ll hear about it in time. 

Maybe we will. But nothing says whoever finds the break has to tell anyone, or tell everyone at the same time. Someone with bad intentions has no reason to publish at all. 

I’m not saying anyone is sitting on a secret RSA break. There’s no public evidence of that, and “they probably have one” isn’t an argument. 

Rotation Is Not Migration

“We rotate our certificates every 90 days, so we’re good.” 

That’s great hygiene. But renewing a certificate or rotating a key doesn’t necessarily change the math underneath it. As a result, a new key of the same type may leave the weakness in place. 

Switching algorithms isn’t your server’s call alone, either. Every client, device, and partner system has to support the replacement. You also need to stop trusting signatures an attacker could forge with the compromised algorithm. 

Rotating a key is routine maintenance. Changing what your systems are willing to trust is a migration, and it’s a much bigger job. 

The Firmware Problem

Attackers can record traffic now and decrypt it later if its cryptographic protection fails, and migrating tomorrow won’t un-steal anything they already have. 

The scenario that bugs me more is signatures. Devices check signatures before installing firmware. If an attacker can forge a trusted vendor’s signature, malware can look just as legitimate as a real update. Even the update meant to replace the vulnerable algorithm may rely on the same broken protection. 

Recovery depends on the device’s design. Some support changing verification keys or using another trusted recovery path. But others have hardware constraints. The way out could mean a backup trust path you set up years ago (you did that, right?), someone physically touching every box, or buying new hardware. 

If your answer to a broken algorithm is a six-year appliance refresh, that’s a dependency you can find today instead of during an incident. 

So What Do We Actually Do?

You don’t have to predict anything. You need to know how long it would take to replace a broken algorithm in the services you can’t live without. Here’s where I’d start: 

  1. Pick a handful of services that really matter. Remote access, identity, software delivery, device management, backups. Don’t try to inventory the whole company on day one. Otherwise, you’ll end up with a spreadsheet nobody trusts. 
  1. Figure out what’s really running. Which algorithm does which job, what depends on it, how long the data has to stay secret, and how you’d change it. A config setting? A vendor release? A forklift? A scan report is a decent start, but it isn’t an inventory. 
  1. Fix what’s already weak. Old protocols, 1024-bit RSA keys, and vulnerable libraries don’t need a future breakthrough to justify replacing them. 
  1. Pilot hybrid post-quantum key exchange where it’s supported. Then check what actually gets negotiated at every hop. Upgrading the connection to your load balancer does nothing for the connection behind it. 
  1. Give signing its own plan. Certificates, code and firmware signing, HSMs, secure boot. A strong key exchange won’t save you if someone can impersonate the endpoint. 
  1. Make vendors show their work. What’s fixed in hardware? What’s shipping today versus “on the roadmap”? Also, can you turn off a vulnerable algorithm without locking yourself out of recovery? 
  1. Run the fire drill. Pick one critical service and pretend its algorithm was just credibly broken. Replace it in a test environment, prove the old algorithm gets rejected, and time each step. Then budget for whatever stopped you from finishing. 

If you only do one thing on this list, do the last one. You’ll know which services you can switch in a day, which are waiting on a vendor, and which need a forklift. That’s the number you take to leadership. 

NIST finalized IR 8547 on October 27, 2026, with accelerated deadlines of 2030 for key establishment and 2031 for digital signatures in specified high-value and high-impact federal systems. The broader federal migration target remains 2035. Those dates aren’t a prediction of when anything breaks, and they definitely aren’t permission to start in 2034. 

The Part You Can Control

Post-quantum crypto won’t make you immune to every future discovery, either. HAWK was still a candidate, and finding its weakness before deployment is exactly what the review process is supposed to do. Being able to change again still matters. 

Nobody can tell you when, or whether, the math under your encryption will give way. Not me, and not the AI labs. Even so, you can find out how long it would take to swap it out. I’d bet the number is bigger than you think. 

Y2K had a countdown. This one doesn’t, so you’ll have to start without being told. 

Not Sure How Long Your Migration Would Take?

Figuring out where your crypto lives and what’s standing in the way of replacing it is exactly the kind of messy problem we like. If you want help mapping it out, let’s talk.