Did AI Just Break RSA? What DeepMind's Paper Actually Says
On November 5, 2026, Google DeepMind released a working paper titled "Fan Factories: Multiple Number Fields in Factorization Factories and Batch NFS". The headlines exploded: "AI breaks RSA", "cryptocurrency apocalypse", "end of encryption". The truth: the paper improves the efficiency constant of an integer-factoring algorithm, and states explicitly that it does not threaten 2048-bit RSA.
This is not the end of cryptography. It is another milestone in AI-assisted mathematics.
1. Headlines vs Primary Source: Point by Point
| Media claim | What the paper actually says | Evidence level |
|---|---|---|
| "AI broke RSA" | The paper improves the constant c of the number field sieve (NFS) from 1.9019 to 1.7929 | A (Abstract) |
| "RSA is no longer safe" | "Our results do not threaten deployed cryptosystems today that are based on 2048-bit and larger RSA." | A (page 1) |
| "Autonomous AI breakthrough" | "The results were obtained with a two-agent AI harness: one model generated the proofs, and a second model verified them." | A (Abstract) |
| "1024-bit RSA cracked" | The paper discusses theoretical cost reductions for 1024-bit integers, a key size deprecated since 2013 | A (Section 4) |
Verdict: the headlines are clickbait; the paper itself is solid number theory.
2. What the Paper Actually Did
2.1 The number field sieve
The number field sieve (NFS) is the fastest known classical algorithm for factoring large integers, and therefore the standard way to attack RSA moduli. Its heuristic running time is usually written as L(1/3, c): the exponent of e is c times the cube root of ln N, times the two-thirds power of ln ln N.
The constant c is the entire scoreboard. Lower c by even a few percent, and the running time drops by an exponential factor - and the bigger N is, the more dramatic the saving.
2.2 A constant untouched for 33 years
| Algorithm | Constant c | Year |
|---|---|---|
| Standard GNFS | 1.923 | 1993 |
| Coppersmith's MNFS | 1.9019 | 1993 |
| DeepMind's fan factory | 1.7929 | 2026 |
From 1993 to 2026, nobody improved this constant for 33 years.
2.3 The new method: fan factory
The paper's idea:
- Give each target integer its own fan of number fields (a family of algebraic polynomials), instead of a single polynomial
- Share one expensive precomputation of the rational side across all targets
- With a precomputation budget of roughly one GNFS run, the per-target constant drops to 1.7929
Size of the improvement:
| Metric | Value |
|---|---|
| Constant | from 1.9019 down to 1.7929, a reduction of 0.109 |
| At 1024 bits | after one precomputation, each further integer costs about 1/58 of a full GNFS run |
| Versus MNFS | about 30x cheaper per integer |
2.4 The real novelty is the method
The highlight is not "RSA broken" but how the result was produced:
| Feature | Description |
|---|---|
| Two-agent AI harness | one model generated the proofs, a second model verified them |
| Machine-checked in Lean 4 | the algebraic identities behind the closed forms are formally verified |
| Closed-form trade-off | for every precomputation budget, the paper gives the smallest achievable per-integer cost |
This is a textbook case of AI-assisted mathematics: the AI did not "crack" anything - it helped discover an optimization path that human number theorists had missed for three decades.
3. Why the Media Exploded: Three Events Muddled Into One
In late 2026, three separate developments around integer factoring landed within weeks of each other, and coverage blended them into a single "AI breaks RSA" story:
| Event | Date | Nature | AI involved? |
|---|---|---|---|
| DeepMind Fan Factories paper | 2026-11-05 | Algorithmic improvement (constant c) | Yes - AI-assisted research |
| RSA-260 (863 digits) factored | 2026-09 | RSA Factoring Challenge record | Unverified |
| Quantum resource estimate, eprint 2026/2665 | 2026-10-30 | Quantum route (Shor's algorithm) | No - unrelated to AI |
3.1 RSA-260: a challenge record, not an AI break
- In September 2026, RSA-260 (an 863-digit semiprime from the famous RSA Factoring Challenge) was reportedly factored, succeeding RSA-250 (829 digits) as the largest broken.
- There is no verified evidence that AI played a special role.
- 863 decimal digits is far smaller than the 2048-bit (617-digit) keys deployed today.
3.2 The quantum line: eprint 2026/2665
- Titled "Shor's algorithm compiled on cat qubits", it estimates that factoring RSA-2048 would take about 22,593 cat qubits running for roughly 3.14 days.
- This is a quantum computing resource estimate, using Shor's algorithm.
- It has nothing to do with AI. No such quantum computer exists today.
3.3 How the confusion happened
| Headline pattern | What it conflates |
|---|---|
| "AI breaks RSA" | Recasts a constant-factor improvement as a break |
| "RSA-260 falls to AI" | Attaches the challenge record to the AI story without evidence |
| "Quantum + AI: the end of RSA" | Merges a quantum resource estimate with an AI math result |
In reality: three independent events. Only the DeepMind paper involves AI, and that same paper explicitly says 2048-bit RSA is not threatened.
4. Practical Security Impact: Nearly Zero
4.1 When fan factories actually help
| Condition | Why it matters |
|---|---|
| Many targets of the same size | The precomputation amortizes only across a batch |
| Huge precomputation budget | Costs roughly one full GNFS run up front |
| Small-ish keys (about 1024 bits) | At 2048 bits and above, the gain shrinks sharply |
4.2 1024-bit RSA has been dead since 2013
| Year | Milestone |
|---|---|
| 2013 | NIST deprecated 1024-bit RSA; minimum 2048 bits recommended |
| Today | Deployed systems use 2048-bit or 4096-bit keys |
Making already-broken 1024-bit keys cheaper to break changes nothing real.
4.3 The paper's own security statement
"Our results do not threaten deployed cryptosystems today that are based on 2048-bit and larger RSA."
The authors also note that the main significance is how quickly AI-assisted research can move a field that usually advances slowly - a comment on the pace of mathematics, not on your bank's TLS certificate.
5. Why Lowering c Is Worth So Much
Key intuition: c is a multiplier inside the exponent, not an additive saving.
| Step | Explanation |
|---|---|
| The constant looks small | 1.9019 to 1.7929 is a drop of only 0.109, roughly 5.7 percent |
| The leverage is exponential | because c sits in the exponent of a subexponential function, the saving grows explosively with N |
| At concrete sizes | at 1024 bits the paper computes about 1/58 of a GNFS run per additional integer, 30x better than MNFS |
And the amortization structure matters even more:
| Cost item | Amount |
|---|---|
| Precomputation (rational-side sieve) | one GNFS run, paid once |
| Each subsequent integer | about 1/58 of a GNFS run |
The precomputation is a one-time, astronomically large cost that nobody wants to pay; but once paid, the marginal cost of every additional key collapses. That is exactly what the word "factory" means in the title - turning one-off costs into amortized batch costs.
This also explains the paper's batch theorem: if you factor M integers together without precomputation (batch NFS), the per-integer exponent improves by at most 0.018, an order of magnitude smaller than the precomputation mode. The batch advantage comes almost entirely from amortizing the shared precomputation.
6. The Real Signal: AI-Assisted Mathematics Hits a Hard Problem
6.1 An AI result on a problem ignored for 33 years
| Significance | Detail |
|---|---|
| 33-year gap | a constant nobody improved since 1993, now improved |
| Novel construction | a fan of number fields per target, not a single polynomial |
| Machine-verified | core algebraic identities checked in Lean 4 |
6.2 Part of a bigger wave
This paper belongs to the same movement as several 2026 events:
| Event | Date | Content |
|---|---|---|
| OpenAI's math manuscripts | 2026-10 | 721 manuscripts on GitHub from an unreleased internal model; 373 result families; about 42 percent of top-line results formalized |
| Anthropic formal-math repo | 2026-11 | three Lean 4 projects published |
| Leiden Declaration | 2026-11 | mathematicians warning against the "serious misalignment" of AI in mathematics |
The common thread: AI is generating mathematics at unprecedented speed, and the mathematical community is reacting at unprecedented force.
6.3 Assessment (marked as analysis)
| Question | Assessment |
|---|---|
| Is this "breaking RSA"? | No. It is AI-assisted mathematics with a narrow, honestly-stated security footnote |
| Is cryptography in danger? | Not from this paper. 2048-bit RSA is not threatened; quantum computing is a separate, slower clock |
| What should we watch? | The pace: AI-assisted research just moved a subfield that had been frozen for 33 years |
For cryptography:
| Timeframe | Impact |
|---|---|
| Short term (5-10 years) | Zero practical impact on deployed RSA |
| Medium term (10-20 years) | Migration to post-quantum cryptography continues; driven by quantum estimates, not by this paper |
| Long term (20+ years) | AI-assisted mathematics may keep shaving constants; whether it can ever reach 2048-bit practicality is unknown and, on current asymptotics, unlikely |
For AI research:
| Trend | Note |
|---|---|
| From toys to hard problems | AI math is moving from contest problems to open research questions untouched for decades |
| A template emerges | generator-verifier agent pairs plus Lean 4 formalization may become the standard workflow |
| Community pushback | the Leiden Declaration shows trust, not capability, is the bottleneck ahead |
7. What to Watch Next
| Item | Why |
|---|---|
| The full version of the paper | current version is a working paper; the full version should carry all proofs and the Lean source |
| Independent reproduction | the algebra is machine-checked; the algorithmic implementation still needs third-party validation |
| Details of the AI harness | which models, what failure rates - this determines whether the method transfers to other fields |
| Smaller keys | below 1024 bits the gains could be larger, but those keys are already deprecated |
One final sober note: even if 1024-bit factoring became 58 times cheaper, the practical payoff for an attacker is tiny - real keys are 2048 bits or larger, where the benefit of fan factories is far smaller. This is a win for AI-assisted mathematics, and a rounding error for cryptanalysis.
8. Sources
| Source | Status |
|---|---|
| DeepMind working paper, Fan Factories (PDF, 2026-11-05) | verified directly |
| Hacker News threads on the paper and RSA challenge | verified directly |
| arXiv 2608.21986, AI Grinding for Fun and Cryptanalysis | abstract verified |
| eprint 2026/2665, Shor on cat qubits | title verified only |
| Bruce Schneier, The Future of AI Cryptanalysis (Oct 2026) | existence confirmed, content not retrieved |
Rigor note: RSA-260 factoring details and the Schneier essay could not be fully verified from primary sources at the time of writing; both are marked accordingly in the text. All constant values and quoted security statements come directly from the DeepMind paper.
9. One-Sentence Summary
"AI breaks RSA" is clickbait: DeepMind used a two-agent AI harness to improve the number field sieve constant from 1.9019 to 1.7929 after 33 years of no progress - the paper itself states that 2048-bit RSA is not threatened. The real story is that AI-assisted mathematics just moved a field that had been frozen since 1993.