Google DeepMind says its AI agents found cheaper ways to attack old 1,024-bit encryption keys

Google DeepMind, Google's AI research lab, says AI agents working with its researchers found two cheaper ways to attack RSA encryption keys that are 1,024 bits long, according to two working papers the lab released Nov. 6.  The papers say the results do not threaten systems that use keys of 2,048 bits or more, today's standard.

What RSA is

RSA is a long-used method computers rely on to scramble data and to create digital signatures, codes that prove a file or message came from its sender and was not changed.  Its safety rests on how hard it is to factor, or break down, a huge number into the two prime numbers multiplied to make it.  Longer keys, measured in bits, make that much harder.  The National Institute of Standards and Technology, the federal agency that sets U.S. security standards, says keys shorter than 2,048 bits must not be used to create new signatures, though 1,024-bit keys may still be used to check old ones.

What the agents did

The first paper looks at an attacker who can briefly get a device to sign data of their choosing, then later forge signatures alone.  Researchers at UC San Diego and France's Inria research institute showed this year that such an attack on a 1,024-bit key works, using a hardware security module, a locked-down device that holds secret keys.  Their job took about five months.

Redoing it with the same method, two cooperating AI agents, one planning and checking and one writing and running code, needed 391 years of single-processor computing time for the first forged result, 3.5 times less than before.  Spread across up to 300,000 processors, it finished in 3.2 days.  A human researcher sent about 20 short messages, mostly approvals, the paper says.  DeepMind says the work ran in a sealed-off test setup on a key that had already been broken, without touching any live systems.

The second paper describes breaking many 1,024-bit keys in a batch.  After an up-front cost about equal to breaking one key, each extra key would cost roughly 1/58 as much as usual, the paper estimates.  That method exists only on paper so far.

Why it matters

The authors say the results sharpen what experts know about how well older signature standards hold up, and show how quickly AI-assisted research can move a field that usually advances slowly.  Co-authors include Nadia Heninger of UC San Diego and Inria, Emmanuel Thomé of Inria and Chirag Pabbaraju, who is also affiliated with Stanford University.  The papers have not been published in a peer-reviewed journal.

Source:  DeepMind paper on forging signatures; DeepMind paper on batch factoring; UC San Diego and Inria paper; DeepMind announcement; NIST key-length guidance