DeepMind AI Agents Find Two Weaknesses in Legacy RSA-1024
3 min readGoogle DeepMind said on November 6 that AI agents working under the direction of human cryptanalysts uncovered two new RSA-1024 weaknesses, and that the work took weeks rather than the years such results usually demand. The company stressed that modern RSA-2048 keys are not affected. The announcement still lands as a warning shot for anyone running old infrastructure that never rotated its keys.
Why RSA-1024 weaknesses still matter
RSA is the public-key system behind much of the internet’s trust, from certificates to code signing. Keys of 1,024 bits were deprecated for new use more than a decade ago, and public certificate authorities stopped issuing them around 2014. Legacy keys of that size survive anyway in embedded firmware signing, old internal PKI certificates, forgotten SSH keys and aging VPN appliances. Breaking one 1,024-bit key is already known to be feasible with enough compute, so what matters is how cheap it gets.
What DeepMind found
In a post on X, DeepMind wrote: “Guided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems.” Coverage from OfficeChai names cryptanalysts Nadia Heninger of UC San Diego and Inria and Emmanuel Thomé of Inria as the experts who steered the agents.
The first result is a signature forgery algorithm. By DeepMind’s theoretical estimates it improves on the best known forgery method by roughly 1,600 times for 1,024-bit keys in the common case where the attacker controls only a small slice of the signed data, such as a 32-byte hash. The team reproduced the attack in a sandbox against a benchmark key that academics had already broken, and did so with about 3.5 times less CPU time. The second result is a batch factoring algorithm with better asymptotic complexity than any previous method. It cuts the cost of factoring each additional 1,024-bit key by about 30 times. ByteIota reports that a paper titled “Forging 1024-bit RSA signatures in nearly SNFS time” has been posted to the IACR ePrint archive.
Why it matters
Pushmeet Kohli, DeepMind’s vice president of science, wrote that “we must not assume the mathematical ‘hardness’ of cryptographic schemes” and argued that AI can help defenders improve standards well ahead of future threats. The deeper point is speed. Cryptanalysis advances normally arrive over years of specialist effort. If agents can compress that into weeks, the comfortable margins baked into today’s key sizes and migration timelines shrink.
This is not a live attack and no production certificate has been forged with it. It is a research result with a sandbox demonstration. The practical takeaway is an audit: find any 1,024-bit RSA key still in service and retire it. Watch for the full paper and for whether the same agent approach gets pointed at other legacy schemes.