Google DeepMind Says Its AI Agents Found Two Major Weaknesses in Legacy 1024 Bit RSA

Context

Google DeepMind shared research on using AI agents, guided by humans, to stress test encryption. Its VP of Science Pushmeet Kohli explained the work in a thread. The weaknesses concern legacy 1024 bit RSA systems, a key size already considered outdated.

Deep dive

Google DeepMind said on Friday that it is sharing new research showing how AI agents can accelerate the science of stress testing encryption. Guided by human researchers, its models uncovered two major weaknesses in legacy 1024 bit RSA cryptosystems, according to the lab.

RSA at 1024 bits has been considered too weak for new systems for years, but it remains in older hardware and software. Research that finds concrete weaknesses in those deployments helps security teams decide what to replace first.

DeepMind framed the work as human guided, meaning researchers directed the agents rather than the models working alone. The announcement came a day after OpenAI introduced GPT 6 Cyber for security research, which shows labs racing to show their models doing real security work.

The posts did not claim a break of modern RSA key sizes. Details of the two weaknesses are in Pushmeet Kohli's thread and the linked research.

What to expect next

Watch for outside cryptographers to review the findings and for vendors still using 1024 bit RSA to respond.

The broader question is how quickly AI driven cryptanalysis moves from legacy systems to current standards, and how labs handle disclosure when their models find real flaws. Expect more labs to publish similar results as AI security work becomes a main way to show what new models can do.