RSA-896 Factored by AI: When Claude Broke a 30-Year-Old Cryptographic Challenge

On September 19, 2026, a post appeared on Hacker News with a deceptively simple title:

RSA-896

The post, authored by someone using the handle saweis, claimed that they had factored RSA-896 — a 270-digit (896-bit) semiprime from RSA Laboratories' famous list of challenge numbers — using Claude. And they provided the factors.

If verified, this is not just a curiosity. It is a seismic event for cryptography, for AI capabilities, and for every system that still relies on 896-bit RSA keys.

What Is RSA-896 and Why Does It Matter?

RSA Laboratories published a series of challenge numbers in 1991, ranging from 100 digits to 617 digits (RSA-2048). The goal was to track the real-world difficulty of factoring large semiprimes — the mathematical problem that underpins RSA encryption. Each number is the product of two large primes, and the challenge is to find those primes.

RSA-896 is a 270-digit number, weighing in at 896 bits. To put that in perspective, 1024-bit RSA was deprecated years ago, and 2048-bit is the current minimum standard. But 896 bits is still no joke — factoring it requires either an impractical amount of classical compute or a breakthrough in algorithmic approach.

Until now, the smallest unfactored RSA challenge numbers were RSA-896, RSA-1024, and RSA-2048. RSA-896 has stood unbroken for over 30 years.

How Was It Done?

According to the post, the factorization was achieved using Claude — Anthropic's AI assistant. The author published both prime factors:

p = 636606729769440499166579950236036751749912014371509557713570027508971809534551913252252094954941974952859310861988904737359709200557919

q = 647218161102195448058768698177623951380616936266986989243011933572862870905830904361851542450154852431416136790787107595965374752513489

The details of the interaction remain sparse. Did Claude perform the factorization directly through chain-of-thought reasoning? Did it write and execute code that implemented a factoring algorithm like the General Number Field Sieve (GNFS)? Did it discover a novel approach? The post does not say, and the cryptography community is already scrambling to verify the result.

What we do know is that modern AI models, particularly those with extended reasoning capabilities, have become surprisingly effective at mathematical problem-solving. Models like Claude, GPT-6, and Gemini have demonstrated the ability to:

• Write and optimize complex mathematical algorithms
• Identify mathematical structure that humans might miss
• Orchestrate multi-step computational pipelines
• Apply creative problem-solving approaches across domains

The Verification Problem

Here is the thing about cryptographic claims: they are only as good as their verification. The RSA challenge numbers have a formal submission process through RSA Laboratories (now owned by Dell Technologies), and independent verification by multiple cryptography groups is essential.

Fortunately, verification is straightforward — multiply the two published primes together and check if the product equals RSA-896. The math is deterministic. If p × q = RSA-896, the factorization is correct. The real question is whether Claude actually produced these factors or whether the human author used traditional methods and attributed the work to AI.

The Hacker News comment thread is already filled with skepticism. Some users point out that GNFS implementations like CADO-NFS could feasibly factor a 896-bit number with sufficient compute — perhaps a few hundred to a few thousand core-hours on modern hardware. The question is whether Claude contributed something algorithmically novel or whether it simply wrote code that executed a known algorithm efficiently.

Why This Matters Beyond the Math

Regardless of the exact mechanism, the implications are significant:

• AI as a cryptographic tool — If AI models can effectively orchestrate complex mathematical computations, they become a force multiplier for cryptanalysis. A single researcher with a capable AI assistant may now accomplish what once required a team of specialists.
• The erosion of key size assumptions — The cryptography community has long used key size as a proxy for security. If AI can shrink the effective compute gap, those assumptions need revisiting. 2048-bit RSA may not be as safe as we think for as long as we think.
• Democratization of cryptanalysis — Advanced mathematical attack tools that once required deep expertise may soon be accessible to anyone with an API key. This is both exciting and terrifying.
• Post-quantum urgency — We already knew quantum computers would eventually break RSA. If AI can accelerate classical attacks, the transition to post-quantum cryptography becomes even more urgent.

The Bigger Picture: AI and Mathematics

This result, if confirmed, fits into a broader pattern. Over the past two years, AI models have been making steady progress on mathematical problems that were once considered the exclusive domain of human mathematicians. From proving theorems to discovering counterexamples to solving competition problems, AI is increasingly demonstrating genuine mathematical reasoning — not just pattern matching, but creative problem-solving.

The RSA-896 factorization may turn out to be a case where the AI simply wrote efficient code to execute a known algorithm. But even that scenario is remarkable. It means that the barrier to mounting a serious cryptanalytic attack has just dropped from 'you need a team of cryptographic researchers and a supercomputer' to 'you need an AI subscription and some cloud compute.'

What Should You Do?

For most organizations, the practical advice remains unchanged but bears repeating:

• Migrate to at least 3072-bit RSA or, preferably, elliptic curve cryptography
• Begin planning your post-quantum cryptography migration now — NIST has already published standards
• Audit your systems for legacy keys that may be using sub-1024-bit RSA
• Treat AI as both a tool and a threat in your security model

The Bottom Line

Whether Claude factored RSA-896 through novel reasoning or by writing efficient code for a known algorithm, the message is clear: the computational landscape is shifting. AI is no longer just a tool for writing emails and generating images. It is becoming a serious player in mathematics and cryptography, and the implications for security, privacy, and infrastructure are profound.

The RSA challenge numbers have stood for 35 years as a benchmark for what is computationally feasible. If AI can now move that benchmark, everything we thought we knew about key sizes and security margins needs to be reconsidered.

We will be watching the verification process closely. If confirmed, September 19, 2026 may go down as the day AI broke one of cryptography's oldest standing records.