Factoring RSA-896 using Claude and GPU Clusters

On September 19, 2026, the RSA-896 challenge number was successfully factored. The effort was led by an Anthropic employee, madars, who utilized the Claude AI model to handle the technical porting and orchestration of the computation required for the factorization.

Implementation and Compute Resources

The factorization of RSA-896 was achieved by porting the CADO-NFS (General Number Field Sieve) implementation to run on GPUs. Claude was used to perform the porting of the CADO-NFS code and to orchestrate a fleet of GPUs utilizing scavenged idle capacity within a data center.

Key technical specifications of the operation include:

  • Hardware: A maximum of 2,048 GPUs were utilized.
  • Duration: The process ran for approximately 10 days.
  • Compute Cost: The total effort amounted to roughly 30 GPU-years of computation.
  • Software: The project relied on the existing General Number Field Sieve (GNFS) algorithm and the CADO-NFS software suite.

Security Implications and Algorithmic Status

This achievement does not represent a breakthrough in cryptographic algorithms or a new threat to currently deployed RSA keys. The factorization was performed using known, existing algorithms that remain exponential in complexity.

As the author of the project noted:

"The credit belongs first to the people who built the number field sieve and CADO-NFS over several decades, and to the teams who set the earlier records. This run used their algorithm and much of their code."

Community Discussion and Analysis

Technical discussions surrounding the event highlighted several points regarding the efficiency of compute and the role of AI in the process:

  • Role of AI: Commenters noted that the AI was not used to discover a new mathematical method for factoring, but rather as a tool for software engineering (porting code) and infrastructure orchestration.
  • Compute Utilization: The use of "scavenged idle capacity" suggests that the operation utilized resources that were already paid for and reserved, making the cost of the run effectively zero in terms of marginal financial expenditure.
  • Practical Vulnerabilities: Some users pointed out that smaller RSA keys, such as 768-bit keys still used by some services like Instagram for DKIM, may now be trivial to factor using similar GPU-accelerated weekend projects.

Sources

Related