Google DeepMind has published new research showing that AI agents, working under human supervision, can dramatically accelerate cryptanalysis — the discipline of stress-testing encryption systems to find their breaking points. The team says its agents uncovered two significant weaknesses in legacy 1024-bit RSA cryptosystems, demonstrating that these older systems can be compromised with far less computing power than previously assumed.
The research was led by Dj Dvijotham and Abhradeep Thakurta, working alongside Nadia Heninger of UC San Diego and Inria, and Emmanuel Thomé of Inria. It lands at a moment when AI labs are increasingly turning agentic systems loose on open research problems rather than narrow benchmark tasks — the same instinct that produced DeepMind’s AI co-mathematician, which recently set a new high score on the FrontierMath Tier 4 benchmark and helped an Oxford mathematician crack a decades-old open problem.
Why RSA-1024, and why it’s not a reason to panic
RSA encryption rests on a simple asymmetry: multiplying two large prime numbers is trivial, but factoring that product back into its original primes is extraordinarily hard. That “hardness” is what keeps encrypted communications, digital signatures, and much of the internet’s security infrastructure intact.
DeepMind is careful to note that the results apply only to obsolete 1024-bit RSA systems. Modern encryption, which almost universally relies on 2048-bit keys or larger, is unaffected. RSA-1024 has long been considered retired and unsafe for real-world use, which is exactly why researchers use it as a sandbox: a system complex enough to be meaningful, but safe to attack without touching anything live.
What makes the finding notable isn’t that a retired system was broken — it’s how quickly the AI agents got there. Historically, cryptanalytic breakthroughs of this kind take years, sometimes decades, giving defenders time to assume that hardness margins shift only gradually. DeepMind’s agents compressed that timeline substantially, which is the real headline: AI can now probe cryptographic assumptions at a pace the field hasn’t had to plan around before.
The first weakness: padding doesn’t protect as much as assumed
The first result concerns digital signatures — the mechanism that lets a computer verify a message, software update, or document genuinely comes from its claimed source. Before signing, many standards add “padding,” extra filler data, to round a message out to a fixed-size block. Cryptographers always knew padding helped prevent forgery, but not precisely how much protection it offered once an attacker could feed chosen inputs into a signing device.
DeepMind’s agents found a sharper answer: protection depends on how much of the signed block an attacker controls. Once an attacker controls roughly half the block or more, padding stops adding meaningful protection — forging a signature becomes nearly as easy as if there were no padding at all. Working entirely in a sandboxed simulation against an already-solved 1024-bit benchmark key from an earlier academic study, the agents reproduced that result using 3.5 times less CPU time, and at under a thousandth of the computing power it would take to break the underlying key itself.
Even in tougher cases — where an attacker controls less of the block, such as with a short 32-byte digital fingerprint — the agents discovered a new algorithm that, by DeepMind’s theoretical estimates, improves on the best previously known method by roughly 1,600x for 1024-bit keys.
The second weakness: factoring many keys at once, cheaper
The second result tackles a different angle: factoring many RSA keys of the same bit length at once. Because RSA keys typically share a standard size — 1024 or 2048 bits — an attacker targeting many keys can, in theory, spread one large upfront computation across all of them, lowering the average cost per key. DeepMind’s agents discovered a new algorithm with better asymptotic complexity than any previously known batch method, cutting the average cost of factoring each additional 1024-bit key by about 30x.
Why DeepMind is publishing this
The company frames the release as a contribution to the broader cryptanalysis community rather than a warning shot. Hardness assumptions have always been stress-tested by researchers using retired systems as safe proving grounds, and DeepMind argues this work simply shows that AI agents can now do that stress-testing far faster than human teams alone — which means cryptographic hardness can no longer be treated as a fixed, slowly-eroding assumption. The intent is to give defenders better tools to measure cryptographic strength proactively, rather than waiting years for the next hand-crafted breakthrough.
Part of a pattern: AI finding what mathematicians missed
This isn’t the first time DeepMind’s agents have turned up results that had eluded human experts for decades. Its AlphaEvolve system, for instance, discovered a new algorithm for multiplying 4×4 complex-valued matrices using just 48 scalar multiplications, improving on Strassen’s 1969 approach for the first time in 56 years — a record mathematicians had tried and failed to beat since before most of today’s AI researchers were born. That same evolutionary coding approach has since been pointed at data-center scheduling, chip design, and even the training pipeline for DeepMind’s own Gemini models.
The cryptanalysis results fit the same mold: take a problem the field assumed was well understood, let an AI agent search far more of the solution space than a human team realistically could, and see what was hiding in the gaps. It’s the same philosophy behind the AI co-mathematician’s run at FrontierMath Tier 4, where the system’s edge came less from raw model size and more from giving agents a persistent, reviewable workspace to try, fail, and retry ideas over days rather than minutes.