DeepMind's AI agents found two new weaknesses in old 1024-bit RSA (2048-bit is safe)
Quick answer
Google DeepMind published two working papers, with cryptographers Nadia Heninger (UC San Diego and Inria) and Emmanuel Thomé (Inria) among the authors, in which human-guided AI agents improved known attacks on legacy 1024-bit RSA. One makes forging certain old-style signatures cheaper; the other makes factoring many 1024-bit keys at once cheaper. Both papers state plainly: “Our results don’t threaten deployed cryptosystems today that are based on 2048-bit and larger RSA” (signature paper, PDF; batch-factoring paper, PDF).
What the AI agents actually did
- A real computation, 3.5x cheaper. The team reran a 1024-bit signature-forgery problem another group had already solved. Same algorithm and settings, but the first result took 391 CPU core-years instead of 1,380, which is 3.5 times less. The paper says the saving came mainly from better code, not better hardware.
- Run almost entirely by agents. Two cooperating AI agents, a supervisor and an executor, designed, ran and debugged it. The human researcher sent about twenty short messages, most of them approvals. The agents wrote about 33,500 lines of new code and ran 336 distributed experiments, finishing in 3.2 days on up to 300,000 CPU cores (the earlier run took about five months, mostly because it had less parallel hardware).
- A new shortcut for old signatures. For the PKCS#1 v1.5 signing format, the paper shows the format stops adding protection once an attacker controls more than half the signed block (or a 64-byte digest at 1024 bits).
- Cheaper bulk factoring, on paper. The second paper’s “fan factories” method estimates that, after a setup costing about one factorization, each additional 1024-bit number costs about 1/58 of a normal factorization, 30 times less than the best comparable method since 1993. That is a theoretical estimate; no one has built it yet. Its algebra was machine-checked in Lean 4, with one AI model writing proofs and a second checking them.
Why it matters
- The headline is less about RSA and more about speed: AI agents carried a major cryptography computation from code to result in days, with humans mostly approving.
- The authors say the main significance is how quickly AI-assisted research “can move a field that usually advances slowly.”
- If you still have 1024-bit RSA keys anywhere (old devices, old certificates, long-lived servers), this is one more reason to retire them.
What it is not
- Not a break of 2048-bit or larger RSA.
- Not peer-reviewed yet. Both are working papers.
- Not an attack on any real system. The computation reran a research problem that had already been solved.
FAQ
- Is my bank or HTTPS connection at risk? Not from this. Both papers say 2048-bit and larger RSA is not threatened.
- What should I do? Most people: nothing. If you run servers or devices, check for leftover 1024-bit RSA keys and replace them.
- Did AI do this alone? No. The papers describe human-guided agents, and the researchers chose the problems and approved the steps.