Google DeepMind says AI agents found two weaknesses in legacy 1024-bit RSA within weeks; modern keys unaffected

Google DeepMind said on November 6 that AI agents working under human guidance found two weaknesses in legacy 1024-bit RSA encryption within weeks: a cheaper way to forge digital signatures and a cheaper way to crack many keys at once. The team stressed that modern 2048-bit and larger keys are not affected; the two working papers have not been peer reviewed. (Photo: file photo of the Google DeepMind logo.)

The two weaknesses: signature forgery and batch cracking

The first weakness concerns digital signatures: when an attacker controls about half of the signed block, signature padding offers little protection. DeepMind said a forgery reproduced in a sandbox took about a third of the CPU time of the original method, and it estimates a new method could be about 1,600 times faster in theory for harder cases such as 32-byte fingerprints. The second weakness, dubbed a “fan factory”, is a new batch-factoring algorithm that could cut the average cost of cracking each additional 1024-bit key by about 30 times — an estimate only; no such machine has been built.

Measured results and estimates are not the same

The only result actually run was a re-run of a known 1024-bit attack: the original team spent 1,380 CPU core-years; DeepMind’s agents did it in 391 core-years, in 3.2 days using up to 300,000 processor cores. The saving came mostly from better code, not a new algorithm. The 1,600x and 30x figures are theoretical estimates. The work was led by DeepMind’s Dj Dvijotham and Abhradeep Thakurta, with Nadia Heninger of UC San Diego and Emmanuel Thomé of Inria; the two working papers were published on November 5 and 6.

Modern keys are safe; old devices remain the risk

DeepMind said explicitly that RSA keys of 2048 bits and above are not affected, so everyday users have nothing to worry about. 1024-bit RSA was already considered obsolete: the US National Institute of Standards and Technology has disallowed it for new signatures since 2013. But some old devices and signed firmware still use 1024-bit keys, and the phase-out is incomplete. The real significance is speed — work that once took human cryptographers months or years was done by AI agents in weeks — though the papers still await peer review and independent replication.