Google DeepMind AI Agents Cut the Cost of Attacking Legacy RSA-1024

Google DeepMind says AI agents supervised by a human researcher significantly reduced the cost of two known attacks against legacy 1024-bit RSA keys. The work does not threaten modern 2048-bit RSA, but it illustrates how AI can accelerate the engineering needed to turn difficult cryptographic research into working attacks.

In one experiment, the agents optimized a previously demonstrated RSA-1024 signature-forgery technique from an estimated 1,380 CPU core-years to 391 CPU core-years, a 3.5-fold reduction. The test used a sandboxed key that had already been solved and required roughly four billion signing queries. It did not recover the private key.

DeepMind also said its agents proposed a theoretical batch-factoring improvement that could make each additional 1024-bit key about 30 times cheaper once a large computation was underway. That method was not implemented, so it remains a research result rather than a demonstrated operational capability.

Justin Beals, CEO and cofounder of Strike Graph, said: “Two AI agents designed, ran and debugged the forgery computation in under four days, with about 20 messages from a human researcher, mostly approvals. The math still took 391 CPU core-years, so nobody is forging signatures on a laptop.”

The practical risk falls on organizations that still have 1024-bit RSA hidden in older appliances, firmware, code-signing systems or vendor infrastructure. DeepMind said 2048-bit and larger RSA keys remain secure against the reported work.

Beals recommended an immediate inventory: “Build a cryptographic inventory that includes your vendors and your HSM configurations.” Security teams should prioritize replacing 1024-bit keys and verify that suppliers are doing the same.

Sources: Google DeepMind | SecurityWeek analysis | Strike Graph