Here's a headline I did not expect to write on a Sunday morning: an engineer at Anthropic says he factored RSA-896, one of the most famous unsolved challenge numbers in cryptography, using Claude. The run finished on September 19, and the result checks out — anyone can multiply the two published prime factors together and get the original number back. So no, this isn't a lab claim you have to take on faith.
01.Wait, what is RSA-896?
Quick refresher for those of us who slept through cryptography class. Most of the internet's security — your banking, your WhatsApp, your UPI transactions — rests on a simple asymmetry: multiplying two huge prime numbers is easy, but taking the result and figuring out which two primes produced it is brutally hard. That's RSA, named after its three inventors.
In the 1990s, RSA Security published a list of "challenge numbers" to test exactly how hard this is. RSA-896 is 896 bits long — about 270 decimal digits, a number that would fill most of a printed page. Weis says the two primes he found are each 448 bits, or roughly 135 digits. Before this, the largest publicly factored number of this type was the 829-bit RSA-250, which took a research team around 2,700 CPU-years back in 2020.
02.How Claude actually did it
The interesting part isn't the math — it's the engineering. The method is the General Number Field Sieve, an algorithm from the early 90s that the open-source CADO-NFS package implements. What's new is who did the work: Weis had Claude port CADO-NFS so it runs on GPUs, and then Claude orchestrated a fleet of them, scavenging idle capacity across the company's clusters.
At peak, the run used up to 2,048 GPUs at once for about 10 days — roughly 30 GPU-years of total computing. To put that in perspective: one GPU running flat out for a year is one GPU-year. Spread across thousands of machines, a job that would have taken three decades on a single chip finished in ten days. When Weis asked Claude if it had a message for the public, it pointed at the humans: "The credit belongs first to the people who built the number field sieve and CADO-NFS over several decades." Honestly, that's the most Indian-heartwarming part of the whole story — the AI giving credit to its elders.
03.No, your passwords are not in danger
Before your WhatsApp groups start panicking, let's be clear about what this does not mean. There is no new mathematics here — Claude did not discover a shortcut, and no new factoring algorithm exists. The problem is still hard, and it gets much harder as numbers grow. Ten days on 2,048 GPUs for 896 bits does not translate into cracking the 2,048-bit keys actually used in practice; those are astronomically more difficult. Weis himself says this poses no new threat to deployed keys. Your net-banking login is fine. Breathe.
04.The real story: agents running big projects
Here's the thing nobody's talking about enough: the significance is the process, not the cryptography. Porting highly-tuned, decades-old scientific software to GPUs and then coordinating a fleet of thousands of chips across scavenged idle capacity is exactly the kind of messy, multi-week engineering project that normally needs a specialist team. Here, one engineer plus an AI agent did it.
And it's not even the first time this month. On September 3, Eric Lu at Cognition — the company behind the Devin coding agent — factored RSA-260, a 260-digit challenge number, using a swarm of Devin agents running a modified GPU version of the same CADO-NFS code, for roughly $400,000 in compute. Two AI-assisted factoring records in under three weeks, from rival camps. If you've been reading our blog, this fits a pattern we covered in detail earlier — from Claude now helping build the next Claude, to GPT-6 Astra cracking a decade-old open math problem on FrontierMath.
05.Why compute costs decided this story
Notice what actually made these records possible: not smarter math, but cheaper orchestration. Thirty GPU-years sounds free when it's "idle capacity," but if you're paying Indian cloud rates for that compute, the difference between a smart approach and a wasteful one is the difference between a ₹50 lakh bill and a ₹5 crore one. The skill that's becoming valuable isn't knowing number theory — it's knowing how to direct AI agents so they don't burn your budget on dead ends. That's exactly why we put a full cost-optimization chapter in our GPT-6 Astra — Zero to Hero Guide, and honestly, it might be the most practically useful chapter in the book.
Related reads
- Claude hacked three companies in Anthropic's safety tests — the model's autonomous streak, from the other side.
- Claude Code finally supports AGENTS.md — more proof of agents running real engineering work.
Key Takeaways
- Anthropic engineer Steve Weis factored RSA-896 (270 digits) with Claude's help; the result is publicly verifiable
- Claude ported CADO-NFS to GPUs and ran up to 2,048 GPUs for ~10 days — about 30 GPU-years
- No new math, no threat to real-world 2,048-bit keys — this is an engineering milestone, not a security breach
- Cognition's RSA-260 result earlier in September used a similar playbook with Devin agents
- The emerging skill: directing AI agents through large projects without burning your compute budget
Frequently Asked Questions
Did Claude break RSA encryption?
No new cryptography was discovered. Anthropic engineer Carrick Weis had Claude port the open-source CADO-NFS package to GPUs and orchestrate a fleet of up to 2,048 GPUs for about 10 days — roughly 30 GPU-years — to factor RSA-896, a 270-digit challenge number, using the decades-old General Number Field Sieve algorithm.
Is my banking or WhatsApp encryption at risk after this?
No. The 2,048-bit keys actually used in practice are astronomically harder to factor than RSA-896, and Weis himself says this poses no new threat to deployed keys. The previous record, RSA-250, took a research team about 2,700 CPU-years in 2020 — this was an efficiency story, not a breakthrough in breaking encryption.
Why does this matter if no new math was discovered?
The significance is the process: one engineer plus an AI agent completed a messy, multi-week specialist engineering project — porting decades-old scientific software to GPUs and coordinating thousands of chips across scavenged idle capacity — that would normally need a full team. It was also the second AI-assisted factoring record in under three weeks, after Cognition's Eric Lu factored RSA-260 with Devin agents for roughly $400,000 in compute.
Related Reads
- OpenAI puts 1 in 4 engineers on AI defense duty — the other side of the coin: AI agents hunting security holes instead of creating them
- Were the AI hacks oversold? — what security insiders actually said after the scare
What This Means for You
If you're a student or developer in India, this is a signal worth reading carefully. The people who ran these record-breaking projects weren't doing anything a well-guided engineer with strong agent skills couldn't attempt — they knew how to break a huge task down, let the AI do the porting and orchestration, and verify the result. That's a learnable skill, and it's getting more valuable every month. If you want to build it, start with the fundamentals: prompting, agent patterns, and cost control. Our GPT-6 Astra — Zero to Hero Guide covers all three from an absolute-beginner level, and it's a lot cheaper than 30 GPU-years.
And if GPT-6 Astra itself is still new to you, our GPT-6 Astra explainer is the best place to start — it covers what the model does, what it costs, and how to stay safe while using it.