Google DeepMind: Human-Guided AI Agents Speed Cryptanalysis; Two Weaknesses in Legacy 1024-bit RSA; 2048-bit+ Remains Safe

Core Background & Industry Pain Points

Modern encryption rests on computational hardness: multiplying large primes is easy; factoring the product is not. Cryptographers have long used retired RSA-1024 as a safe proving ground, but major advances historically took years. On 2026-11-06, Google DeepMind reported that human-supervised AI agents can probe these hardness assumptions faster, lowering the computational budget needed against legacy 1024-bit systems—while stating that widely deployed 2048-bit and larger protocols remain unaffected.

Architecture Highlights & Internals

The work is led by Dj Dvijotham and Abhradeep Thakurta, with Nadia Heninger (UC San Diego / Inria) and Emmanuel Thomé (Inria). One line targets signature padding: if an attacker can feed chosen inputs to a signing device and control roughly half the signed block, padding adds no protection and forgery collapses to the no-padding case. The second line targets batch factoring of many same-size moduli: agents found an algorithm with better asymptotic complexity, cutting the incremental cost after a large shared precomputation. All experiments were sandboxed; no live systems were touched.

Authoritative Benchmarks & Measured Scores

Figures below are DeepMind / collaborator-reported, not independently reproduced here. In sandbox replay on an already-solved 1024-bit academic benchmark key, the agent-engineered implementation used about 3.5× less CPU than prior work and under one-thousandth the compute of a full key break. For short fingerprints (e.g. 32 bytes), theoretical estimates put a new algorithm about 1600× better than the previous best at 1024 bits. For batch factoring many same-size keys, average cost per additional 1024-bit key is about 30× lower than prior batch methods.

Developer Hands-on Guide

Read Pushmeet Kohli's X Article and the two GCS technical PDFs (Faster Affine Root for signatures; Fan Factories / Batch NFS for factoring). Inventory any remaining 1024-bit RSA certs, HSMs, or blind-signing oracles; migrate production to 2048+ and plan post-quantum paths. When building agent security tooling, include controlled signing oracles and padding fraction in the threat model—without overstating today's 2048-bit risk.