Google DeepMind has used AI agents working under human supervision to find two new weaknesses in legacy 1024-bit RSA encryption, one of the oldest locks on the internet. The agents reproduced a signature forgery using 3.5 times less computing power than the team that first performed it, and produced a new method for attacking many old keys at once. The findings do not threaten modern encryption: DeepMind is explicit that 2048-bit RSA and larger keys are unaffected, and no live system or live key was touched.

Summary: Announced on 6 November 2026, the research shows AI can now do in weeks what once took cryptographers months of dedicated work. The results apply only to obsolete 1024-bit RSA, a key size the US standards body stopped accepting for new signatures in 2013. Your bank account, messages and software updates are not at risk from these specific findings.

Key takeaways

  • DeepMind AI agents found two weaknesses in legacy 1024-bit RSA: easier signature forgery and cheaper batch factoring.
  • The measured result is a 1024-bit forgery reproduced in 391 CPU core-years instead of 1,380, a 3.5 times saving.
  • The headline 1,600 times figure is a theoretical estimate for one narrow case, not a measured attack.
  • Modern 2048-bit and larger RSA keys are unaffected, and the work ran entirely in an isolated sandbox.
  • The real story is speed: AI is changing how fast cryptographic assumptions can be stress-tested.

What DeepMind announced

Google DeepMind announced the research on 6 November 2026 in an article by Pushmeet Kohli, its vice president of science, titled “Advancing the frontier of cryptanalysis”. The work was led by DeepMind researchers Dj Dvijotham and Abhradeep Thakurta, working with cryptographers Nadia Heninger of the University of California San Diego and Inria, and Emmanuel Thomé of Inria. It is described in two working papers published on 5 and 6 November 2026, “Modular Roots from a Restricted Oracle” and “Fan Factories”. Neither paper has yet been peer reviewed.

The announcement describes two distinct results, both aimed at the same retired target. The first makes it easier to forge digital signatures on 1024-bit RSA keys. The second makes it cheaper to factor many 1024-bit keys in a batch. Both were produced by AI agents guided by human experts, not by a fully autonomous system, and both were tested in an isolated environment against a key size chosen precisely because attacking it endangers nobody.

The two DeepMind findings at a glance, with the source of each figure
Finding What it does Figure Measured or estimated
Signature forgery Forges signatures when the attacker controls about half the signed block or more 391 vs 1,380 CPU core-years Measured: 3.5 times less CPU
Signature forgery, small control New algorithm for cases such as a 32-byte fingerprint About 1,600 times better Theoretical estimate
Batch factoring Cuts the average cost of factoring each additional 1024-bit key About 30 times cheaper Theoretical estimate; nothing built
Scope 1024-bit RSA only 2048-bit and larger unaffected DeepMind statement, in both papers

What RSA is, and why 1024 bits stopped being enough

RSA is a public-key encryption system named after its inventors, Rivest, Shamir and Adleman, and it has underpinned internet security since the late 1970s. Its security rests on a simple bet: multiplying two huge prime numbers is easy, but working backwards from the product to find those primes is brutally hard. The product is called the modulus, and factoring it is the key to breaking the lock.

The “bit” number describes the size of that modulus. A 1024-bit key was the standard for much of the 2000s, but computing power kept growing and the safety margin kept shrinking. In 2013, the US National Institute of Standards and Technology stopped accepting 1024-bit RSA for generating new digital signatures. Modern systems almost universally use 2048-bit keys or larger, and the jump from 1024 to 2048 bits is far bigger than it sounds: each extra bit roughly doubles the work an attacker faces, so a 2048-bit key is enormously harder to attack than a 1024-bit one.

Digital signatures are the other half of the story. A signature works like a wax seal on a letter, proving who sent a message and that nobody tampered with it on the way. Before signing, extra preset information called padding is added to the data, and padding has long been understood to make forgery hard. DeepMind’s first finding challenges that assumption in specific conditions.

What the AI agents actually did

The agents’ first task was to reproduce a known result. In September 2026, a team including Heninger and Thomé published a paper showing that 1024-bit RSA signatures could be forged in nearly the time needed for a related, easier problem. That computation consumed about 1,380 CPU core-years spread across five calendar months, using roughly 4.07 billion queries to a signing oracle.

DeepMind’s two-agent harness reproduced that forgery in 391 CPU core-years, about 3.5 times less computing power, and debugged the implementation in under four days. Along the way the agents found and resolved four listed problems affecting the algorithm or the correctness of the result, without needing another pass through the oracle. The algebraic identities and feasibility witnesses behind the result were machine-checked in Lean 4, a proof assistant that verifies mathematical claims mechanically. The result is part of a fast-moving week for AI systems: Tencent released Hy4, a model family aimed at vision and computer use, days earlier.

There is a caveat worth stating plainly. The 1,380 and 391 core-year figures were produced on different hardware, so the comparison is not perfectly like for like. The papers bound that caveat with a same-machine test of the sieving step, which ran 1.8 to 4.4 times faster than the standard CADO-NFS software while finding the same relations, and sieving accounts for about 90 percent of the saving. The forgery also does not recover the private key: it forges a signature without factoring the modulus.

The agents showed that once an attacker controls roughly half of the signed block or more, padding stops adding meaningful protection, and forging a signature becomes nearly as easy as if there were no padding at all. For the harder cases, where the attacker controls only a small slice of the block, such as a 32-byte digital fingerprint, the agents produced a new algorithm that DeepMind estimates improves on the best previously known method by roughly 1,600 times for 1024-bit keys.

The 1,600 times number, read carefully

The 1,600 times figure is the number most headlines carried, and it is the number most easily misunderstood. It is a theoretical estimate, not a measured result. Nobody ran a timed attack that was 1,600 times faster. What DeepMind’s team did was derive a new algorithm and estimate, from the mathematics, that it would improve on the previous best method by about that factor in the narrow case where the attacker controls only a small part of the signed data.

The measured number is the 3.5 times CPU saving on the solved benchmark key. That is still significant: a computation that took a research team five months of calendar time was reproduced in days. But the gap between 3.5 times measured and 1,600 times estimated is the gap between what was demonstrated and what the mathematics suggests could be possible. Both matter, and they should not be conflated.

The same discipline applies to the second finding. The “fan factory” paper proposes a new algorithm for batch factoring: pay a large setup cost once, about the price of one full factoring, and each additional 1024-bit key of the same size then costs roughly 30 times less than the usual multi-target method. The idea gives every target its own family, or “fan”, of number fields, all drawing on the same stored precomputation, which makes mid-sized precomputations worthwhile for the first time. But no fan factory has been built. The 30 times figure is a leading-order estimate from the paper’s analysis, not a measured attack on real keys.

Is your encryption safe?

Yes, for any system using modern key sizes. DeepMind states directly, in bold, in both working papers, that the results do not threaten deployed cryptosystems based on 2048-bit and larger RSA. Interesting Engineering, which covered the announcement, reports that the findings do not threaten online banking, messaging apps, software updates or digital currencies. No live keys were broken, and the experiments ran in simulation on a retired standard.

The people who should pay attention are those still running 1024-bit RSA anywhere. That means old embedded devices, legacy industrial systems, ageing certificate infrastructure and software that has not been updated in a decade or more. For those systems, the message is not new, but it is louder: 1024-bit RSA was already considered weak, NIST ended its acceptance for new signatures thirteen years ago, and the cost of attacking it is now falling faster than expected. Migration to 2048-bit or larger keys, or to modern alternatives, is the fix, and it has been the recommended fix for years.

For everyone else, the practical advice is unchanged and undramatic: keep software updated, use current browsers and messaging apps, and do not reuse passwords. None of those habits is affected by this research, because none of them depends on 1024-bit RSA.

Why this matters beyond RSA

The significance of the announcement is not that an old lock got weaker. Cryptographers already knew 1024-bit RSA was weak. The significance is that AI agents, guided by humans, found new cryptanalytic results in weeks, and that the results were good enough to be written up as papers with machine-checked proofs. DeepMind chose RSA-1024 as a proving ground precisely because it is complex enough to be a meaningful test and retired enough that attacking it harms nobody. The same approach could now be turned on other cryptographic assumptions.

Kohli’s own conclusion points in that direction. In DeepMind’s announcement of the work, he argues that as AI becomes increasingly capable of solving advanced mathematical problems, the field can no longer rely on historical hardness assumptions and must continuously stress-test cryptographic standards. That is a statement about method, not about a single key size: the assumptions that underpin today’s encryption were set when the best attackers were humans with years to spare. AI changes the economics of the attack. The same week, the UK National Cyber Security Centre warned that AI-powered ransomware is highly likely to strike critical infrastructure by 2028, showing that AI is accelerating both attack and defence.

There is also a longer-term context. The cryptography community has spent the last decade preparing for a different threat, quantum computing, and NIST has already standardised post-quantum algorithms such as ML-KEM, ML-DSA and SLH-DSA, which do not rely on integer factorisation at all. DeepMind’s results are classical optimisations, not quantum speedups, and they leave those post-quantum standards untouched. But they reinforce the same lesson the quantum debate taught: cryptographic standards need continuous re-examination, because the tools available to attackers do not stand still. The computing power behind these agents keeps growing, and the same AI data centre build-out is behind Marvell paying 115 billion dollars for optical specialist Lumentum.

The work now faces the usual scientific scrutiny. The papers have not been peer reviewed, outside replication of the 391 core-year computation has not yet happened, and the most striking numbers are estimates awaiting independent verification. That is normal for working papers, and it is not a reason to dismiss them. It is a reason to read the headlines carefully: a theoretical 1,600 times improvement on a retired key size is a research result, not a security emergency.

Frequently asked questions

Has AI broken RSA encryption?

No. DeepMind’s AI agents found two weaknesses in legacy 1024-bit RSA, a key size retired years ago. Modern 2048-bit and larger RSA keys are unaffected, and no live system or live key was attacked. The work ran in an isolated sandbox.

What is 1024-bit RSA, and why is it obsolete?

RSA is a public-key encryption system whose security depends on how hard it is to factor a large number into its prime factors. A 1024-bit key was standard in the 2000s, but computing power has grown so much that NIST stopped accepting it for generating new digital signatures in 2013. Modern systems use 2048-bit keys or larger.

What does the 1,600 times figure actually mean?

It is DeepMind’s theoretical estimate that a new signature-forgery algorithm improves on the previous best method by about 1,600 times for 1024-bit keys, in the case where an attacker controls only a small part of the signed data, such as a 32-byte fingerprint. It was not measured in a timed attack. The measured result was a 3.5 times CPU saving on a solved benchmark key.

Should I do anything differently because of this research?

If you use up-to-date software, browsers and messaging apps, no. If you operate systems that still use 1024-bit RSA certificates or keys, this is another reason to migrate them to 2048-bit or larger keys, a migration NIST has recommended since 2013.

Conclusion

DeepMind’s announcement is best read as a demonstration of a new research capability, not as a warning about the encryption you use every day. AI agents, steered by human cryptographers, reproduced a months-long computation in days, found a sharper answer to a long-standing question about signature padding, and proposed a cheaper way to attack many old keys at once. The target was deliberately obsolete, the modern standard is untouched, and the headline numbers need the careful reading DeepMind itself provides. What has changed is the speed at which cryptographic assumptions can be tested, and that change will outlast this particular result.

Sources and Further Reading

Primary sources

  • Google DeepMind, Pushmeet Kohli, “Advancing the frontier of cryptanalysis”, 6 November 2026
  • Google DeepMind, “Modular Roots from a Restricted Oracle: Faster Affine Roots and a 1024-bit Computation”, working paper, 6 November 2026
  • Google DeepMind, “Fan Factories: Multiple Number Fields in Factorization Factories and Batch NFS”, working paper, 5 November 2026
  • Shea, Haller, Suhl, Heninger and Thomé, “Forging 1024-bit RSA signatures in nearly SNFS time”, IACR ePrint 2026/2131, 20 September 2026 — the earlier paper the DeepMind agents reproduced

Secondary reporting

  • Interesting Engineering, “Google DeepMind finds 1,600x faster way to expose encryption flaws”, 6 November 2026
  • Crypto Briefing, “Google DeepMind’s AI agents expose two weaknesses in legacy 1024-bit RSA”, 6 November 2026
  • AI Frontier Post, “DeepMind’s AI agents found two cracks in legacy 1024-bit RSA, and modern keys are untouched”, 7 November 2026
  • DEV Community, “Google DeepMind: AI agents found two weaknesses in legacy 1024-bit RSA”, 7 November 2026
  • AI News Online, “DeepMind RSA-1024 AI agents explained: what was shown”, 8 November 2026
  • Guruji Corporation, “Two AI Agents Cut a 1,380-Core-Year RSA Attack to 391, and Proved the Maths in Lean 4”, 7 November 2026
  • Google DeepMind on LinkedIn, announcement of the cryptanalysis research, 6 November 2026