TL;DR
Two researchers using two different AI systems broke consecutive RSA factoring records in 16 days, compressing what once took years of specialist effort into weeks and raising fresh urgency around legacy encryption infrastructure.
What happened
- On September 19, 2026, researcher Stephen A. Weis factored RSA-896 (a 270-digit semiprime) with help from Anthropic's Claude, pushing the public record from 862 bits to 896 bits.
- Just 16 days earlier, on September 3, Eric Lu of Cognition used Devin AI agents to factor RSA-260 at 862 bits, setting the prior record.
- Both efforts used the general number field sieve (GNFS), the classical algorithm dominant since the early 1990s. No quantum computing was involved.
- Lu's RSA-260 factoring consumed roughly 4,900 GPU-days at an estimated cost of $400,000. Weis has not published compute costs for RSA-896.
- The previous major milestone, RSA-250 at 829 bits, was factored in 2020 using purely classical methods, making the jump to 896 bits in six years a meaningful acceleration.
Why it matters
- AI is compressing multi-year factoring projects into weeks: what previously required rare number-theory expertise and months of manual parameter tuning, Claude handled as a collaborator navigating computation configuration.
- The bit-difficulty gap is not linear: each additional bit roughly doubles factoring difficulty, so RSA-2048 (the widely deployed standard) remains orders of magnitude harder than RSA-896. Modern, properly configured systems are not immediately threatened.
- Legacy infrastructure is the real exposure: systems still using 1024-bit RSA keys (flagged as insufficient years ago) now face a concrete, accelerating signal to audit and upgrade.
- Two AI companies, Anthropic and Cognition, now hold consecutive credits on public factoring records, signaling that AI-assisted cryptanalysis is becoming a repeatable, competitive capability rather than a one-off stunt.
- The acceleration curve may steepen: if AI tooling keeps compressing the gap between records, the timeline for when larger key sizes come under practical pressure shortens faster than prior projections assumed.
What to watch next
- Whether Weis publishes compute costs for RSA-896, which would let analysts benchmark how much AI optimization actually reduced the GPU-day burden versus Lu's $400,000 RSA-260 run.
- Whether a third record attempt targeting RSA-1024 is announced in the near term, which would be the threshold that directly threatens real deployed infrastructure and would force immediate policy responses.
- How standards bodies and major cloud providers respond: any accelerated guidance from NIST or updated TLS deprecation timelines would confirm that this acceleration is being taken seriously at the infrastructure level.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.