Google DeepMind says AI agents, AI programs that can carry out tasks on their own, helped its researchers find two weaknesses in RSA encryption keys that are 1024 bits long, an older size. The lab announced the work on Friday, Nov. 6, with two technical papers. It says the results affect only systems still using 1024-bit keys, and do not threaten modern systems, which almost all use keys of 2048 bits or more.
RSA is a method computers have relied on for decades to keep data safe. Its safety comes from a fact about numbers: multiplying two large primes is easy, but finding the two primes from their product is far harder. DeepMind says researchers test that idea on retired key sizes such as 1024 bits, which it calls safe proving grounds.
The first result concerns digital signatures, which computers use to check that a software update, document or website really comes from its claimed source. Many signing standards add fixed filler, called padding, before they sign. DeepMind says an attacker may be able to get a signing device to sign inputs of their choosing. Once the attacker's own data makes up about half or more of what gets signed, the lab says, padding adds no protection, and forging a signature gets as easy as if there were no padding.
In a sealed-off test that touched no live systems, the lab says its agents ran an earlier academic attack on a 1024-bit test key from that study, using the same method but 3.5 times less computer time. That is less than a thousandth of the computing power needed to break the key itself. The paper says two cooperating AI agents designed, ran and debugged the computation in under four days. The computing work added up to about 391 years for a single processor core.
The second result is about breaking many keys of the same size, where, in theory, an attacker can spread one large up-front computation across all of them. DeepMind says its agents found a new method that cuts the average cost of each extra 1024-bit key by about 30 times compared with earlier methods, once the attacker has done an up-front computation about as costly as breaking one key. The paper reaches that figure by calculation, not by breaking real keys. It says one AI model wrote the mathematical proofs behind it while a second checked them.
Both papers are labeled working papers, meaning drafts. The team included cryptographers Nadia Heninger, of UC San Diego and the research institute Inria, and Emmanuel Thomé, of Inria. DeepMind argues that advances like these used to take years or decades, and that encryption's safety margins can no longer be treated as fixed. How much the second method would save on real hardware is not yet known.