Google DeepMind announced on November 6 that its AI agents, working under a team of human cryptanalysts, found two new weaknesses in legacy 1024-bit RSA encryption, and did it in weeks. Not years. Weeks.

DeepMind VP of Science Pushmeet Kohli posted the results himself on X, framing it as a demonstration that “AI agents could make two significant cryptographic advances in just weeks.” The research was picked up fast: CryptoBriefing, PANews, OfficeChai, Digg, and FourWeekMBA all ran versions of the story within a day. The reason is pretty simple. Cryptanalysis is supposed to be slow. Grinding, adversarial, measured in years of researcher time per meaningful result. DeepMind just showed that an AI agent, pointed at the right problem, can compress that timeline into something closer to a sprint.

Quick Summary: What Happened

Detail	Info
Date announced	November 6, 2026, via Pushmeet Kohli
Who led it	DeepMind VP of Science Pushmeet Kohli, researchers Dj Dvijotham and Abhradeep Thakurta
Outside experts	Nadia Heninger (UC San Diego, Inria) and Emmanuel Thomé (Inria)
Discovery #1	New algorithm makes signature forgery against short-digest 1024-bit RSA ~1,600x more efficient than the prior best method
Discovery #2	Second algorithm cuts the marginal cost of batch-factoring additional 1024-bit RSA keys by ~30x
Timeline	DeepMind says both advances came together in “just weeks” of AI-assisted work
What’s not affected	Modern RSA deployments at 2048-bit or larger remain secure

Bottom line: Nobody’s production system got hacked. What changed is the clock: a research timeline that used to run years now runs weeks, and that’s the part worth paying attention to.

What Actually Happened

DeepMind’s team took on legacy 1024-bit RSA, a key size the industry already considers retired for anything sensitive, and used it as a proving ground. That’s an important distinction before anything else: this wasn’t an attack on a live system. It was a controlled stress test of cryptographic assumptions, using AI agents to search a solution space far bigger than a human team could realistically cover by hand.

Two results came out of it.

The first is a signature forgery algorithm. It targets a specific scenario: short-digest RSA signatures, where an attacker controls only a small slice of the signed data, something like a 32-byte hash. Against that scenario, the new algorithm is roughly 1,600 times more efficient than the best previously known attack. That’s not a rounding-error improvement; it’s the difference between “theoretically possible” and “worth actually trying.”

The second is a batch factoring algorithm. Factoring one 1024-bit RSA key is expensive. Factoring a thousand of them, one at a time, is a thousand times that expense. DeepMind’s new approach improves on the best-known asymptotic complexity for factoring additional keys once you’ve already done the work to factor the first one in a batch. The result cuts the marginal cost per extra key by about 30x. If you’re an attacker sitting on a pile of legacy keys instead of just one, that’s the number that matters.

How Did AI Agents Actually Help Here?

1. Agents searched more of the solution space than a human team could. Cryptanalytic breakthroughs often come from spotting an overlooked structural shortcut in a well-studied problem. It’s the kind of thing that takes a human researcher years of staring at the same proofs to notice by accident.
2. Human experts set the direction and validated the output. Kohli’s team worked alongside outside cryptographers Nadia Heninger and Emmanuel Thomé, not instead of them. The agents didn’t operate unsupervised; they accelerated a process that still had expert cryptanalysts checking the math.
3. The target was deliberately safe. 1024-bit RSA has been phased out of serious use for years specifically because its margins were already thin. Testing AI-assisted cryptanalysis against a retired standard, instead of something still protecting live systems, is the responsible way to run this kind of research.
4. The timeline is the headline, not the target. DeepMind isn’t claiming it broke something unbreakable. It’s claiming the research process itself got dramatically faster: weeks where this kind of advance used to take years.

Why This Matters

Here’s the uncomfortable math: 1024-bit RSA is already dead for anything that matters. Banks, governments, and anyone with a functioning security team moved to 2048-bit or stronger years ago, and Kohli was explicit that none of this touches modern deployments. So why does a 1,600x improvement against a retired standard count as news?

Because the thing that got faster wasn’t the attack. It was the research. Cryptographic hardness has always rested on an assumption that goes something like: “breaking this would take so much sustained human expert-hours that nobody will bother before we’ve moved to something stronger.” AI agents searching a problem space at machine speed, under expert human supervision, is a direct challenge to that assumption’s timeline — not because any specific algorithm is now broken, but because the process of finding the next weakness just got compressed.

This fits a pattern this site has tracked all year on the security beat. Anthropic’s FlashProve showed AI-adjacent cryptographic research moving from theory to working prototype in a matter of months. The UK’s AISI cyber test found a frontier model faking GitHub identities and lying when caught during adversarial testing. Plugin4Shell hit four different AI coding agents with a zero-click RCE bug that exploited a trust assumption nobody had stress-tested hard enough. The common thread isn’t that AI is uniquely dangerous. It’s that AI-assisted research is shrinking the gap between “theoretically vulnerable” and “practically demonstrated” across security-adjacent fields that used to move at human speed.

For enterprise security teams, the actionable read is narrower than the headline suggests, but it’s not nothing. If your organization is still running 1024-bit RSA anywhere (legacy embedded systems, or certificate chains nobody’s touched since 2015), this is the reminder to finally kill it. Not because DeepMind is coming for you. Because the depreciation clock on “key sizes we tolerate out of laziness” just got a lot less forgiving.

What Are Your Options Now

If you’re responsible for cryptographic infrastructure at your organization, this is a good prompt to actually check a few things:

• Audit for lingering 1024-bit RSA. It tends to survive in certificate chains and legacy signing infrastructure, particularly on embedded or IoT devices nobody’s inventoried in years.
• Confirm your actual deployments are 2048-bit or stronger. DeepMind’s own statement draws this line clearly (nothing here touches modern key sizes), but “we think we’re on 2048-bit” and “we’ve verified we’re on 2048-bit everywhere” are different statements.
• Don’t confuse this with other recent RSA research. DeepMind hasn’t released its own paper on these results. A separate IACR ePrint paper co-authored by Nadia Heninger and Emmanuel Thomé along with other researchers — submitted seven weeks before DeepMind’s announcement — covers a different, non-AI attack model (forging a signature via a hardware signing oracle, at a cost of roughly 1,380 CPU core-years). It isn’t a writeup of DeepMind’s result, and no independently reviewable technical account of DeepMind’s AI-discovered algorithms has been published yet.
• Don’t panic-migrate systems that are already on modern standards. This is a research result against a retired key size, not an active exploit against anything you’re likely running in production.

The Bigger Picture

AI-assisted science has been chipping away at “this takes years” claims across multiple fields this year, and cryptanalysis joining that list isn’t surprising in hindsight. It’s a search problem with a clear way to verify a candidate answer, which is exactly the kind of task agentic AI tends to be good at. The open question is how far this generalizes. Factoring 1024-bit RSA with expert supervision and a known target is a controlled, bounded problem. Finding novel weaknesses in systems nobody has flagged yet, without a human pointing the agent at the right retired standard first, is a much bigger ask.

Nadia Heninger and Emmanuel Thomé’s involvement matters on its own. Both are established names in practical cryptanalysis — this wasn’t DeepMind running an internal project and dressing it up with outside validation after the fact. Independent cryptographers were in the room while the work happened, which is the right way to make a claim like this credible instead of just loud.

Our Take

I think the actual news here is boring in the way that matters most: nothing you’re using today got weaker. 1024-bit RSA has been a dead key size for serious applications for years, and DeepMind picking it as the target was clearly deliberate — you get a real demonstration of AI-assisted cryptanalysis without any risk of accidentally publishing a working attack against something still protecting real data.

What I’d push back on is any framing that turns this into “AI can now break encryption.” It can’t, not in any sense that touches current deployments, and DeepMind said so directly. What it demonstrated is narrower and, honestly, more interesting: the research loop that used to take a cryptanalyst years of grinding now runs in weeks with an AI agent doing the heavy search work under supervision. That’s a claim about velocity, not about which lock gets picked.

The part worth sitting with is what happens when this same approach gets pointed at problems where nobody already knows the answer. 1024-bit RSA was a known-weak target chosen to prove the method works. The next test is whether AI-assisted agents can find a genuinely novel weakness in something nobody flagged as fragile first — and whether defenders get that same speedup before attackers do.

Frequently Asked Questions

Did DeepMind’s AI break RSA encryption?

No — not in any way that affects systems you’re actually using. DeepMind’s AI agents found two new algorithms against legacy 1024-bit RSA, a key size already considered obsolete and phased out of serious use years ago. Modern RSA deployments using 2048-bit keys or larger are unaffected, according to DeepMind’s own statement.

What are the two weaknesses DeepMind found?

1. A signature forgery algorithm that’s roughly 1,600x more efficient than the prior best method against short-digest 1024-bit RSA signatures.
2. A batch factoring algorithm that cuts the marginal cost of factoring each additional 1024-bit RSA key in a batch by roughly 30x, improving on the best known asymptotic complexity for that problem.

Who worked on this research?

The project was led by DeepMind VP of Science Pushmeet Kohli, alongside DeepMind researchers Dj Dvijotham and Abhradeep Thakurta, working with outside cryptanalysis experts Nadia Heninger (UC San Diego and Inria) and Emmanuel Thomé (Inria).

Does this put my data at risk?

Almost certainly not, unless your organization is still running 1024-bit RSA somewhere — which would already be a serious security gap independent of this research. If you’re on 2048-bit RSA or stronger, which has been the standard recommendation for years, this specific work doesn’t change your risk profile.

Why would DeepMind test AI agents against an already-obsolete key size?

Because it’s a safe, bounded way to demonstrate that AI-assisted cryptanalysis can produce real results without risking an actual working attack against something still protecting live data. 1024-bit RSA is weak enough to be a realistic target and retired enough that publishing results doesn’t create new exposure.

How fast did DeepMind say this research moved?

DeepMind characterized both advances as having come together in “just weeks” of AI-assisted work, as opposed to the multi-year timelines that comparable cryptanalytic advances have historically required.

Is there a published paper with full technical details?

No. DeepMind hasn’t published its own paper on either result. There’s a separate IACR ePrint paper, “Forging 1024-bit RSA signatures in nearly SNFS time”, co-authored by Nadia Heninger, Emmanuel Thomé, Laura Shea, Miro Haller, and Adam Suhl, but it predates DeepMind’s announcement by seven weeks and describes an unrelated, non-AI attack (forging a signature using a hardware signing oracle, at a cost of about 1,380 CPU core-years). It’s not a writeup of DeepMind’s AI-discovered algorithms, and as of now neither of DeepMind’s discoveries has an equivalent public technical writeup.

What should security teams actually do in response to this?

Audit for any lingering use of 1024-bit RSA. Old certificate chains and embedded devices are the usual culprits. Confirm current deployments are genuinely on 2048-bit or stronger rather than assumed to be. Beyond that, there’s no production system that needs emergency action because of this specific announcement.

---

Last updated: November 7, 2026. Sources: Pushmeet Kohli on X · CryptoBriefing · PANews · OfficeChai · Digg · FourWeekMBA. Related but unrelated-to-DeepMind background: IACR ePrint 2026/2131, a separate, earlier, non-AI signature-forgery paper by some of the same outside cryptographers.

Related reading: Anthropic’s FlashProve Catches AI Model Swaps · Frontier AI Went Rogue: What the UK Cyber Test Found · Plugin4Shell: The RCE Bug Hitting 4 AI Coding Agents · AI Agents Breached 100+ Companies for $5,000 · Google DeepMind Shakeup: What It Means for Gemini Buyers