Back to wire
Research·Article·Confirmed

RSA-896 falls as Stephen Weis reports a Claude-assisted 2,048-GPU run

Stephen A. Weis published two prime factors for the 896-bit RSA challenge number. Their product is directly checkable; his account that Claude ported CADO-NFS and orchestrated up to 2,048 GPUs remains first-hand evidence rather than an independently reproduced method report.

Published 20 Sept 2026, 03:31 · Updated 21 Sept 2026, 09:43

The factorisation itself is directly checkable

Weis published RSA-896 in full alongside two 448-bit factors. Multiplying those factors reproduces the 896-bit, 270-decimal-digit challenge number exactly, and standard primality testing identifies both published factors as prime. That makes the mathematical result independently checkable without taking an AI system or a social-media post on trust.

The result moves the public RSA Factoring Challenge frontier beyond the 862-bit RSA-260 result reported earlier in September. That comparison is about a benchmark number designed to track practical factoring capability; it does not mean an 896-bit key was taken from a live system.

The Claude and GPU account is first-hand evidence

Weis says he had Claude port CADO-NFS to GPUs and then use the resulting tooling to orchestrate scavenged idle capacity. His published account puts the peak fleet at 2,048 GPUs, the wall-clock run at about 10 days and aggregate compute at roughly 30 GPU-years.

Those engineering details are attributable to Weis. A comparable public technical report, code diff, job log or independent reproduction of the GPU port has not yet been published, so the factorisation can be confirmed more strongly than the exact path used to obtain it.

The algorithmic boundary matters

Weis explicitly says the run introduced no meaningful new factoring algorithm. CADO-NFS implements the General Number Field Sieve, the established classical method used for large general-purpose integer factorisation. On his account, Claude contributed software engineering and distributed-compute orchestration around that existing method rather than discovering a cryptanalytic shortcut.

That distinction is the useful signal for agentic engineering: mature scientific software can gain practical reach when porting, tuning and large-fleet operation become cheaper to organise. It does not by itself establish that the same automation will reproduce the result on another cluster or generalise to other computational mathematics workloads.

What it does and does not say about RSA security

RSA-896 remains far smaller than the 2,048-bit RSA keys commonly deployed today. Neither Weis nor the published factorisation supplies evidence that RSA-2048 can be factored with comparable resources, and Weis states that the result creates no new threat to deployed keys.

Weis has separately argued that 1,024-bit RSA is within reach of actors with data-centre-scale GPU fleets. That is an extrapolation from current classical factoring work, not a demonstrated RSA-1024 factorisation. Cost, scaling and implementation assumptions would need independent analysis before that claim can be treated as an operational estimate.

Source trail

3 sources · 2 primary · 1 secondary