Google DeepMind said on November 6 that AI agents, working under human supervision, found two new weaknesses in legacy 1024-bit RSA in a matter of weeks. One is a much cheaper way to forge signatures when an attacker controls part of the signed data. The other cuts the cost of factoring many 1024-bit keys at once. According to Interesting Engineering and OfficeChai, the work was led by DeepMind's Dj Dvijotham and Abhradeep Thakurta with cryptographers Nadia Heninger (UC San Diego) and Emmanuel Thomé (Inria). DeepMind VP Pushmeet Kohli announced it on X.
Why it matters: this is new cryptanalysis, not just a faster computer. As reported, the agents came up with new algorithms, and better algorithms shrink a key's safety margin no matter how much hardware an attacker has. DeepMind's own framing, per the coverage, is that old hardness assumptions need continual re-testing. Kohli spoke of "the need to continuously reassess."
The two results
1. Signature padding protects less than assumed. RSA signatures add padding to the data before signing. The researchers report that once an attacker controls roughly half of the signed block or more, padding adds almost no protection. Forging then becomes about as easy as against unpadded RSA.
- Validation ran in a sandbox against an already-solved 1024-bit benchmark key from earlier academic work. The agents' implementation reproduced the known result with 3.5x less CPU time. It also used under one-thousandth of the compute needed to factor that key.
- For the harder case, where the attacker controls only a small piece such as a 32-byte hash, DeepMind estimates its new algorithm is about 1,600x better than the best known method for 1024-bit keys. Both outlets stress this figure is theoretical.
2. Batch factoring gets cheaper. If an attacker wants many RSA keys of the same size, a big upfront computation can be shared across all of them. The agents found a batch method with better asymptotic complexity than any previously known one. As reported, it cuts the average cost of each additional 1024-bit key by about 30x. Neither outlet says it was run at scale.
The human work it builds on
Heninger and Thomé are also co-authors of a September IACR ePrint paper, "Forging 1024-bit RSA signatures in nearly SNFS time." In that paper, a UC San Diego and Inria team actually forged 1024-bit RSA signatures without factoring the key. They used temporary access to a raw signing oracle, a hardware security module they drove through its API. It cost about 1,380 CPU core-years with 2^32 oracle queries, then about 180 core-years per forged signature. The authors estimate that, in that oracle attack model, RSA is 15–30 bits weaker than factoring-based estimates. They add that even 4096-bit RSA falls short of 128-bit security. The coverage we read doesn't say whether that paper supplied DeepMind's benchmark key. Either way, the trend is clear: practical attacks on RSA keep landing below the old estimates.
What it does not mean
- RSA-2048 and larger are not affected, per DeepMind as reported. Banking, messaging and modern TLS are not broken.
- No live system was attacked. The forgery result was shown in simulation against a known key. The 1,600x figure is an estimate, not a demonstration.
- We could not read DeepMind's own write-up or paper, and the outlets don't name the model behind the agents. Treat the exact multipliers as DeepMind's claims until a paper is public.
What to check
1024-bit RSA should already be gone, but it lingers. byteiota's audit checklist points to old embedded and IoT firmware signing, pre-2014 internal PKI, unrotated SSH keys, legacy VPN appliances and HSMs with fixed keys as likely hiding places.
# SSH key size (first number is bits)
ssh-keygen -l -f ~/.ssh/id_rsa.pub
# Certificate or live TLS endpoint
openssl x509 -in cert.pem -noout -text | grep "Public-Key"
openssl s_client -connect host:443 </dev/null 2>/dev/null | openssl x509 -noout -text | grep "Public-Key"
Replace anything reporting 1024 bit. Use Ed25519 for SSH, and ECDSA P-256 or RSA-3072+ for certificates and signing keys. Do fleets first: the batch-factoring result matters most where many same-size keys exist, such as device fleets and certificate pools. And if you are planning a post-quantum migration anyway, this is one more reason not to stretch RSA's lifetime.
Sources
- Interesting Engineering — Google DeepMind finds 1,600x faster way to expose encryption flaws interestingengineering.com
- OfficeChai — Google DeepMind says it has used AI agents to break legacy RSA-1024 encryption faster than anyone thought possible officechai.com
- IACR ePrint 2026/2131 — Forging 1024-bit RSA signatures in nearly SNFS time (Shea, Haller, Suhl, Heninger, Thomé) eprint.iacr.org
- byteiota — DeepMind AI found new RSA-1024 weaknesses: audit now byteiota.com