Claude helps factor RSA-896 as AI-assisted cryptography reaches a new public record

Stephen A. Weis has published the factorization of RSA-896, a 896-bit RSA challenge number, saying that the computation was completed with the assistance of Claude on September 19, 2026. Weis published the original 270-digit integer together with its two factors, allowing the result itself to be checked mathematically rather than accepted as an unverifiable model claim.

Data Studios independently multiplied the two published factors and confirmed that their product is exactly RSA-896. The number has 896 bits and 270 decimal digits, while each published factor is a 448-bit, 135-digit integer.

The result follows Cognition researcher Eric Lu's factorization of RSA-260, completed on September 3 and documented on September 9. That earlier result established a public RSA Factoring Challenge record at 862 bits, using a GPU-optimized implementation of the General Number Field Sieve developed and operated with Cognition's Devin agents. Cognition reported approximately 4,900 GPU-days, or 13.5 GPU-years, of computation and an estimated market-value compute cost of roughly $400,000.

RSA-896 therefore raises the publicly demonstrated threshold from 862 to 896 bits only sixteen days after the previous computation finished.

The result does not mean that Claude discovered a way to break modern RSA encryption, nor does it imply that 2048-bit RSA keys can now be factored economically. Its technical significance lies elsewhere: AI-assisted engineering has now contributed to two successive record-scale classical factorization efforts within weeks, suggesting that frontier models and agents can materially accelerate the implementation, optimization and orchestration of highly specialized scientific computing workloads.

··········

RSA-896 EXTENDS THE PUBLIC FACTORING FRONTIER FROM 862 TO 896 BITS.

The published factors are directly verifiable, while the jump follows almost immediately after Cognition and Devin established the previous RSA challenge record.

RSA challenge numbers were created to provide concrete benchmarks for the practical difficulty of integer factorization. An RSA challenge number is constructed as the product of two large prime numbers. Multiplication is computationally easy; recovering the original factors from the resulting semiprime becomes increasingly expensive as the number grows.

That asymmetry is fundamental to the RSA cryptosystem. Weis's publication supplies both factors of RSA-896, and Data Studios verified the arithmetic relationship p × q = RSA-896 while measuring both factors at 448 bits each.

........

Metric	RSA-896 result
Challenge number	RSA-896
Size	896 bits
Decimal length	270 digits
First factor	448 bits / 135 digits
Second factor	448 bits / 135 digits
Date reported	September 19, 2026
AI system credited	Claude
Previous public RSA challenge record	RSA-260, 862 bits
Increase in bit length	34 bits
Data Studios arithmetic verification	Published factors multiply exactly to RSA-896

........

The 34-bit increase should not be interpreted as a simple 3.9% increase in computational difficulty. Integer-factorization cost does not scale linearly with key length. A relatively small change in the number of bits can require substantially more computation, while improvements in algorithms, implementations, hardware utilization and parallelization can reduce practical cost.

RSA-250, an 829-bit challenge number, had remained the largest publicly solved RSA challenge of this class since 2020. Cognition then moved the frontier to 862 bits with RSA-260, and RSA-896 has now pushed it another 34 bits higher.

··········

CLAUDE'S ROLE APPEARS TO BE ENGINEERING AND COMPUTE ORCHESTRATION, NOT NEW FACTORING MATHEMATICS.

The underlying method belongs to the established Number Field Sieve family, while the AI contribution is associated with adapting and operating the computational pipeline.

CADO-NFS is an open-source implementation of the Number Field Sieve, the standard classical framework used for factoring sufficiently large general integers. Its software already supports distributing parts of the computation across multiple machines.

The RSA-896 result therefore does not require a claim that Claude discovered an unknown mathematical shortcut. The more consequential capability is the ability of an AI system to work around a mature, highly specialized scientific codebase and contribute to the engineering required to make the computation practical at much larger scale.

That pattern closely resembles what Cognition documented for RSA-260. Eric Lu reported that Devin helped produce a GPU-optimized lattice siever, modify multiple stages of CADO-NFS, run experiments, tune parameters, debug failures and orchestrate large-scale cluster jobs. Lu described his own role as setting priorities, establishing benchmarks and redirecting work when the agents moved in unproductive directions, rather than manually implementing every optimization.

Cognition explicitly reported no fundamental algorithmic advance in the RSA-260 work. Instead, the improvement came primarily from GPU performance engineering applied to the existing GNFS pipeline.

The most immediate AI capability demonstrated here is therefore not autonomous mathematical discovery. It is the compression of specialized engineering work: understanding an existing codebase, adapting it to different hardware, running large experimental workloads, coordinating distributed computation and iterating over bottlenecks.

··········

TWO AI-ASSISTED RECORDS IN WEEKS SHOW HOW THE COST OF SPECIALIZED COMPUTATION IS CHANGING.

RSA-260 and RSA-896 provide unusually concrete examples because the final result is objectively verifiable and the engineering workload is substantial.

Cognition reported that its RSA-260 computation consumed approximately 13.5 GPU-years. Public details associated with the RSA-896 work currently provide much less technical documentation than Cognition's extensive RSA-260 report.

That asymmetry in documentation is important. The factorization itself can be verified directly. The precise division of labor between Claude, Weis, existing CADO-NFS code and the compute infrastructure requires more detailed technical reporting before the AI contribution can be quantified with comparable precision.

........

Dimension	RSA-260	RSA-896
Size	862 bits	896 bits
Decimal digits	260	270
Completion	Sep. 3, 2026	Sep. 19, 2026
AI assistance	Devin	Claude
Classical foundation	GNFS / CADO-NFS	Number Field Sieve / CADO-NFS-related workflow
Reported aggregate compute	~13.5 GPU-years	Full technical accounting not yet documented in the primary post
Detailed technical write-up	Yes	Not yet comparable
Result independently checkable	Yes	Yes

........

The rapid succession matters because AI agents are particularly well suited to workloads that contain large amounts of software engineering, experimentation and infrastructure management around an already understood scientific method. Those activities historically required scarce domain specialists to write kernels, manage distributed jobs, inspect failures, tune parameters and repeatedly modify software.

If the same pattern generalizes, the consequence extends beyond cryptanalysis. Computational chemistry, physics simulations, optimization, numerical mathematics and other research domains contain mature algorithms whose practical limits depend heavily on implementation quality and available engineering labor.

··········

RSA-2048 REMAINS IN A COMPLETELY DIFFERENT COMPUTATIONAL REGIME.

A record challenge factorization should not be confused with a practical break of the RSA key sizes used by modern security systems.

Modern RSA deployments commonly use 2048-bit keys, far larger than RSA-896. Cognition's scaling analysis after RSA-260 estimated RSA-2048 to be roughly 91.2 billion times the computational difficulty of RSA-260 under its GNFS scaling model. Those figures are estimates rather than guarantees, but they illustrate why extrapolating directly from an 862- or 896-bit record to 2048-bit keys would be misleading.

The Number Field Sieve remains computationally demanding as integer size grows. CADO-NFS implements a pipeline involving polynomial selection, relation collection through sieving, filtering, large sparse linear algebra and final square-root computations; distributing those stages more effectively changes practical performance without eliminating the underlying scaling problem.

RSA-896 therefore represents a record in classical computation assisted by frontier AI engineering, not a collapse of public-key cryptography.

Its broader implication is more interesting than an exaggerated security headline. Within a matter of weeks, two AI systems from different organizations have been credited with meaningful roles in record-scale factoring projects based on established mathematics. If that progression continues, the near-term effect may be less about discovering instant cryptographic shortcuts and more about making extremely specialized computational research accessible to much smaller human teams.

That is a measurable expansion of what AI agents can do: not merely generate code, but help transform established algorithms into large, optimized and operational scientific systems capable of moving real computational records.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

[datastudios.org]