AI agents accelerated two attacks on legacy RSA without breaking modern keys

Google DeepMind used human guided AI agents to produce two research results on obsolete 1024-bit RSA systems. The authors explicitly say modern protocols using keys of 2048 bits or more remain secure. The work ran in a sandboxed simulation and did not touch live systems.

Agents found faster routes to signature forgery and batch factorization

The first result concerns digital signature padding. When an attacker controls roughly half of the signed block or more, the authors find that padding adds no protection. The agent implementation reproduced an already solved benchmark attack with 3.5 times less CPU time and less than one thousandth of the computing power required to break the key itself.

For a shorter input such as a 32-byte fingerprint, the researchers describe a new algorithm that theoretical estimates put at roughly 1,600 times better than the previous method for 1024-bit keys. A second result improves asymptotic complexity when factoring many numbers of equal size and cuts the average cost of each additional 1024-bit key by about 30 times.

AI shortens the path from mathematical hypothesis to implementation

The important point is not an RSA alarm but the speed of the research cycle. Agents helped explore assumptions, devise algorithms and build a working implementation under cryptographers' supervision.

For defenders, that suggests testing security margins more often. A standard can remain practically secure even when the cost of a known attack class shifts. It is better to discover that shift on retired 1024-bit systems than on infrastructure still protecting live traffic.

A theoretical speedup is not an attack on today's internet

The 1,600-fold figure is a theoretical estimate under specific conditions, while the 30-fold figure concerns average cost across a batch of keys. Neither means common 2048-bit keys have been broken. The team that built the agents is reporting the results, and the technical papers still require scrutiny from the cryptographic community.

Reproduction will show whether agents change cryptanalysis pace

Independent verification of the algorithms, constants and practical costs is the decisive signal. Similar results on other retired systems and a clear account of what agents proposed versus what humans supplied will matter too. Without that boundary, the contribution of agentic research cannot be measured.

Lilith's verdict

Modern RSA remains behind a locked door, but agents have redrawn the map of an old lock. Cryptographers now need to verify the scale.