An Anthropic engineer just used Claude to help crack a famous cryptography homework problem from the 1990s—and the important part is how, not a sudden end to online banking. According to OfficeChai’s Sep 20 report, Steve Weis announced that he factored RSA-896, a 270-digit (896-bit) RSA Security challenge number, with the run finishing on September 19, 2026.
Weis says Claude ported the open-source CADO-NFS implementation of the General Number Field Sieve onto GPUs, then orchestrated a fleet of scavenged idle chips—peaking at about 2,048 GPUs over roughly 10 days, totaling around 30 GPU-years of compute. OfficeChai verified the published prime factors multiply back to RSA-896. Anyone can repeat that check: multiply the two 448-bit primes and confirm you recover the challenge composite.
What RSA-896 actually is
RSA encryption rests on an asymmetry everyone in security already knows: multiplying two huge primes is easy; factoring the product is hard. In the 1990s, RSA Security published a ladder of “challenge numbers” so researchers could measure progress against that hardness assumption. RSA-896 sits at 896 bits—about 270 decimal digits, roughly a page of digits if you printed it. It is a benchmark, not a key your bank issued last Tuesday.
Weis reports the secret factors are each 448 bits (~135 digits). That is exactly the shape these challenges were designed to have. The announcement’s value is historical and operational: another rung on the public ladder fell, and an AI coding stack helped pull it.
Old math, new ops
Nothing here is a novel factoring algorithm. GNFS dates to the 1990s; CADO-NFS is the workhorse package prior record attempts used. What Weis flags as new is the engineering path: adapting a CPU-tuned scientific stack to GPUs and managing thousands of devices on unused capacity—the kind of campaign that normally eats a specialist team’s calendar.
That process story is why the announcement is landing in AI circles as much as crypto-math circles. Models are moving past autocomplete into multi-day projects that include compute logistics, not just code snippets. Porting, scheduling, monitoring, and verifying a sieve run across idle silicon is closer to “staff engineer + SRE” work than to a chat demo.
What it does not mean
Weis and OfficeChai both stress the limits Geeknewz will repeat loudly:
- No new math shortcut — Claude did not invent a faster factoring algorithm.
- Cost still explodes with size — ten days on 2,048 GPUs for 896 bits does not imply practical breaks of 2,048-bit keys used in production TLS and similar systems.
- Deployed keys are not newly at risk — Weis states plainly this is an engineering demonstration, not a reason to panic about bank logins or messaging apps.
In short: cool receipt for agentic coding + GPU scavenging; not a CVE for the internet. Anyone rewriting the headline as “AI breaks RSA” is doing readers a disservice.
Second AI-assisted RSA hit in three weeks
Weis’s result follows a similar Sep 9 publication from Cognition’s Eric Lu, who factored RSA-260 (260 digits / 862 bits) on September 3 using a swarm of Devin agents and a GPU-accelerated CADO-NFS fork on spare Cognition clusters—about 4,900 GPU-days (~13.5 GPU-years) at an estimated ~$400,000. Before these two 2026 runs, OfficeChai notes the largest public comparable was RSA-250 (829 bits) in 2020 at roughly 2,700 CPU-years.
Two different organizations, two challenge numbers, one shared recipe: classical GNFS, GPU acceleration, agent-assisted engineering, scavenged or spare capacity. That repetition is the news. It suggests the workflow is becoming reproducible rather than a one-off stunt.
When Weis asked Claude for a public message, the model credited the decades of humans behind the number field sieve and CADO-NFS. Fair. The industry takeaway still stands: AI systems are now credible co-pilots for large, messy scientific engineering campaigns, including the ugly parts—porting, scheduling, and babysitting fleets.
Geeknewz take
Read past the “AI factors RSA” headline. The news is that Claude helped turn idle AI silicon into a classical cryptanalysis cluster in ten days, weeks after Devin’s team did the same class of work on a slightly smaller challenge. If you care about encryption strength, keep using modern key sizes and ignore the panic bait. If you care about what agents can ship, update your priors about how fast specialist HPC workflows can be revived—and who gets to point them at overnight GPUs.
Source: OfficeChai — Claude Helped Factor RSA-896, A 270-Digit Encryption Challenge, Says Anthropic Engineer (Sep 20, 2026).