Google DeepMind turned two AI agents loose on 1024-bit RSA — the retired encryption standard — and the agents came back with two new ways to pick it. The work, announced Friday by DeepMind vice president of science Pushmeet Kohli, shows AI compressing cryptanalysis from years to weeks. The headline number, 1,600 times faster, is real but theoretical. The measured number — 3.5 times less CPU time — is the one to take seriously.
The first finding targets digital signatures, which verify that software updates are genuine. Signature schemes add preset "padding" before signing, long assumed to block forgery. DeepMind's agents found that once an attacker controls roughly half of the signed block, padding offers virtually no extra protection. To prove it, the agents re-ran a benchmark attack an academic team had already solved, reproducing the result with 3.5 times less CPU time. For the harder case, a 32-byte digest like SHA-256, a new algorithm is theoretically about 1,600 times better than the previous best method — an estimate, not a measured run.
The second finding is a batch-factoring shortcut. Since deployments share key sizes, expensive precomputation can be spread across many keys; DeepMind's "fan factory" algorithm cuts the average cost of each additional 1024-bit key by about 30 times versus the best method known since 1993. No working implementation exists — this one is on paper only.
The limits matter more than the headlines. These results touch only obsolete 1024-bit RSA; modern systems almost universally use 2048-bit keys or larger, and those are unaffected. NIST stopped accepting 1024-bit RSA for new signatures in 2013, and the forgery attack needs direct access to a signing device plus roughly 4 billion oracle calls. Your bank account and messages are not at risk.
But the direction of travel matters. One model supervised, writing the design and auditing results; a second wrote and ran the code. Steered by ~20 short messages from a human, they wrote 33,500 lines of code across 336 experiments. The academics whose work they reproduced spent 1,380 CPU core-years over five months; the agents matched it in 391 core-years, in four days. DeepMind published two working papers for the security community to check. Cryptanalysis is a race between lock-testers and lock-pickers — and the test just got dramatically faster. The locks we use are still holding — the question is how long the assumptions behind them stay still.
— via WebProNews