DeepMind AI Agents Make 1024-Bit RSA Keys Cheaper to Attack

Google DeepMind’s artificial intelligence (AI) agents, working in collaboration with cryptographers, have identified new methods that significantly reduce the computational cost of attacking obsolete 1024-bit RSA cryptographic systems.

The research, conducted with experts from UC San Diego and Inria, reveals that compromising legacy RSA-1024 keys now requires less computational effort than previously understood. However, Pushmeet Kohli, Google DeepMind Vice President and Google Cloud Chief Scientist, emphasised that modern encryption protocols, which almost universally employ 2048-bit or larger keys, remain secure.

The findings detail two primary attack cost reductions. The first involves digital signatures and padding. DeepMind’s AI agents successfully re-ran a known academic forgery attack on a 1024-bit key in a sandboxed environment, cutting the CPU time 3.5-fold. This reduced the initial forgery from an estimated 1,380 CPU core-years to 391 core-years.

This signature forgery method applies to the PKCS#1 v1.5 format, commonly used by the PKCS#11 signing interface in cryptographic devices such as hardware security modules (HSMs). The AI agents designed, executed, and debugged the complex computation in under four days, with minimal human oversight, mostly approvals from human researchers.

While effective, the signature attack does not recover the private key and necessitates approximately 4 billion pre-made signing queries. It is also ineffective against devices that hash messages internally before signing.

The second finding concerns factoring multiple RSA keys of the same size. DeepMind’s agents developed an algorithm that, in theoretical estimates, makes factoring each additional 1024-bit key approximately 30 times cheaper on average than prior methods. This approach assumes a high initial computation cost, comparable to factoring a single key, and has not yet been implemented in practice.

Kohli stated that the research signals a “new era of AI-enabled mathematics and computation,” where cryptographic hardness should not be treated as a fixed assumption. He added that leveraging AI to rigorously test existing cryptographic assumptions is crucial to ensuring the continued security of modern encryption.

Organisations are strongly advised to ensure their systems use current cryptographic standards, specifically RSA keys of 2048 bits or greater, to mitigate such risks.