Factoring RSA 260
A Cognition researcher used an autonomous agent swarm to build a GPU lattice siever and factor RSA-260, setting a new public record for cryptographic challenges.
read it Offers a detailed, practical look at orchestrating autonomous coding agents to solve a complex high-performance computing and cryptography problem.
summary
A researcher at Cognition used an autonomous agent swarm to build a high-performance GPU lattice siever and factor RSA-260, setting a new public record for public RSA challenges. Operating on stranded GPU cluster compute, the agents modified CADO-NFS components to achieve significant speedups over previous methods. The author estimates that scaling this approach could make breaking RSA-1024 feasible for around $30 million, though RSA-2048 remains completely unaffected. This project demonstrates the viability of using AI agents for end-to-end performance engineering and computational mathematics.
key points
- Devin autonomously built a high-performance GPU lattice siever, modifying CADO-NFS to factor the 260-digit RSA challenge using 4,900 GPU-days (~$400k).
- The entire workflow ran on stranded, fragmented single-node GPU compute from LLM training clusters at zero marginal cost.
- The author estimates that hyperscalers could factor RSA-1024 for roughly $30 million, while RSA-2048 remains secure and billions of times harder.
who should read it
Systems engineers, HPC developers, and practitioners interested in autonomous software engineering agents.
discussion warm
consensus
Commenters are heavily impressed by the practical use of autonomous agents for deep performance engineering, while analyzing the real-world costs and implications of GPU-based sieving.
pushback
A few commenters push back on the economic comparisons, arguing that modern CPU improvements make traditional CPU-based factoring much cheaper than stated.
One problem is that the core-years for RSA-250 were in 2017 Skylake Xeon single core terms.
If you can rent a dedicated 96-core Epyc for $1/hour (cheap dedicated host), the combination of IPC improvements (>2x) and core count make 7,010 "skylake core-years" cost only $175k, not $5M. On-demand cloud servers (which cost more than $1, maybe $5/hour) probably make the GPU cheaper, but it's closer than the author says.
It highlights how improvements in CPU core counts and IPC significantly close the cost gap between CPU and GPU factoring methods.
Generated by an LLM from the article and thread. Check the sources before relying on it.