Google's AI Just Found New Cracks in Old Encryption, No Quantum Computer Needed

We usually hear that quantum computers are the big future threat to encryption. Google DeepMind just showed that AI might get there another way. The lab says its AI agents, working with human cryptographers, found two new weaknesses in 1024-bit RSA, an older type of encryption key.

First, the important part: your bank and your phone are fine. 1024-bit RSA was retired for new use more than a decade ago, and larger keys like RSA-2048 aren't affected. No real-world system has been broken. But plenty of old hardware still quietly runs on these keys, including legacy security modules, ageing VPN boxes, older SSH keys, and the firmware signing used in many internet-connected gadgets.

Here's what the agents found. One attack lets someone with temporary access to a signing device forge new signatures without ever stealing the private key. In the toughest version of that scenario, DeepMind estimates its method is about 1,600 times faster than the best previous approach. The second finding makes cracking lots of keys in bulk cheaper. After one big upfront computation, each extra key costs about 30 times less to break than with older methods. The team checked its math with a proof assistant called Lean 4. Both figures are DeepMind's own estimates, and the papers haven't been peer-reviewed yet.

There was also a real-world test. In September, researchers from UC San Diego and France's INRIA forged 1024-bit RSA signatures using about 1,380 CPU core-years over five months. DeepMind says two AI agents repeated that job in under four days using roughly 3.5 times less computing power.

DeepMind's Pushmeet Kohli said the lesson is that we shouldn't assume encryption is as hard to break as we've always believed and that AI can help defenders tighten things up before attackers catch on.

Is AI about to become the biggest threat to old security systems, and should companies be scrambling to replace aging keys now?

Image credits: gagadget