TL;DR
Anthropic engineer Stephen Weis used Claude to factor RSA-896, a 270-digit cryptographic benchmark, pushing the public factoring record to 896 bits just 16 days after the previous record fell, and establishing AI coding agents as the new engine of cryptanalysis progress.
What happened
- On September 19, 2026, Weis published the full prime factorization of RSA-896, a 896-bit semiprime from RSA Security's historic Factoring Challenge, verified by multiplying the two primes to reproduce the original number.
- Claude ported the CADO-NFS codebase from CPU to GPU and orchestrated the compute fleet, handling systems engineering that normally requires a dedicated specialist team over months.
- The run peaked at 2,048 GPUs and consumed roughly 30 GPU-years of work, completed in approximately ten days using scavenged idle compute capacity.
- This is the second consecutive AI-assisted factoring record in 16 days: Cognition's Devin agent broke RSA-260 (829 bits) on September 3, 2026, at an estimated cost of $400,000 and 4,900 GPU-days.
- The frontier moved 67 bits in 16 days, compared to a six-year gap between the 2020 record (RSA-250, 829 bits) and these September 2026 results.
Why it matters
- AI agents have eliminated the specialist bottleneck in large-number factorization: GPU-porting a complex research codebase like CADO-NFS previously demanded months of expert work; Claude reduced it to a one-engineer project.
- Two independent teams, two different AI systems (Devin and Claude) reached consecutive records within weeks, confirming this is a repeatable pattern, not a one-off.
- The cost curve is collapsing: Lu's Devin run achieved GPU sieving at 10 times lower cost than prior CPU-based art, and Weis's run used spare capacity rather than dedicated cluster allocations, lowering the barrier further.
- Deployed infrastructure is not at immediate risk: 2,048-bit RSA and elliptic-curve keys remain astronomically harder to factor than 896-bit numbers, and security experts including Weis have been explicit on this point.
- The strategic implication is forward-looking: if AI agents keep compressing the time and cost to set new records, the timeline for when larger key sizes become relevant to defend accelerates.
What to watch next
- Whether RSA-1024 becomes a near-term target: at 1,024 bits it is the next psychologically and practically significant threshold, still used in some legacy systems.
- Cost and compute trajectories: if scavenged idle GPU capacity can break 896-bit numbers, watch for attempts using purpose-built clusters or next-generation NVIDIA hardware to push further.
- Policy and standards responses: NIST and enterprise security teams may accelerate deprecation timelines for sub-2048-bit RSA if public records keep falling at this pace.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.