DeepMind AI agents expose two weaknesses in legacy 1024-bit RSA
Human-guided agents make legacy-key forgery and batch factoring far cheaper. RSA-2048 is untouched — but it's time to audit your 1024-bit certs.

Copy markdown
Signature forgery, reproduced on the cheap
DeepMind's agents, working under cryptographers Nadia Heninger (UC San Diego/Inria) and Emmanuel Thomé (Inria), reproduced a known 1024-bit RSA signature forgery using 3.5x less CPU than the reference run. For short 32-byte digests, a new algorithm is estimated ~1,600x more efficient than the best prior method — theoretical, but it guts what little protection legacy padding had left.
Batch factoring drops ~30x per key
The second weakness attacks factoring many same-size keys at once: a new algorithm with better asymptotic complexity cuts the marginal cost of each additional 1024-bit key by roughly 30x. If you're sitting on a pile of old 1024-bit keys, the whole pile just got cheaper to pick apart than any single one.
RSA-2048 is not affected
DeepMind is blunt that this touches only obsolete 1024-bit RSA — modern 2048-bit-and-up keys are safe from these specific exploits, and there's no named production victim. It's a research announcement, not a live compromise. The real signal: AI agents compressed weeks of expert cryptanalysis into a repeatable workflow.
What this means for you
If you ship embedded devices, run aging certificate chains, or maintain long-lived infra, inventory your key sizes now. 1024-bit was already deprecated — but deprecated isn't gone. Rotate anything still on 1024-bit before the same AI-assisted probing becomes routine for attackers, not just a DeepMind lab.