Google DeepMind Agents Find Two RSA-1024 Attacks
Google DeepMind’s human-guided agents produced two working papers on legacy 1024-bit RSA: a measured 3.5× cheaper signature-forgery reproduction and theoretical estimates of roughly 1,600× faster restricted forgery and 30× cheaper batch factoring. The authors say 2048-bit-and-larger systems are unaffected.
The headline is bigger than the immediate exploit: DeepMind did not crack a live RSA key, but it showed agentic engineering can compress months of cryptanalytic implementation into days.
- –The measured result reused an existing attack, cutting CPU cost from 1,380 to 391 core-years; the gain came from implementation and massive parallelism, not a fundamentally new factoring algorithm.
- –The 1,600× figure is theoretical and applies to a 32-byte controlled region in a signing interface; it has not been demonstrated at full 1024-bit scale.
- –The “fan factory” method targets batches of same-size RSA keys after shared precomputation, but it remains unimplemented and relies on asymptotic estimates, heuristics, and substantial memory.
- –Security teams should audit legacy RSA-1024 signing APIs and HSMs, especially interfaces accepting attacker-shaped inputs; the work adds urgency to RSA migration and post-quantum planning.
DISCOVERED
1h ago
2026-11-08
PUBLISHED
2h ago
2026-11-08
RELEVANCE
AUTHOR
IntCyberDigest