Google DeepMind agents surface two weaknesses in legacy 1024-bit RSA encryption
On November 6, 2026, Google DeepMind said AI agents working under human supervision had found two weaknesses in legacy 1024-bit RSA. <cite index="22-5">OfficeChai, the original report on the announcement, centered its coverage on how quickly the agents got there.</cite> The first weakness concerns signature forgery. The agents' method, judged only in theory, would be roughly 1,600 times more efficient than the best prior approach for 1024-bit keys when an attacker controls just a fragment of the signed data. <cite index="2-5">The estimate applies to cases such as a 32-byte digital fingerprint.</cite>
The second weakness is a batch-factoring method that makes each further 1024-bit key about 30 times cheaper to factor on average than earlier batch methods. <cite index="9-9">In a sandboxed run, the agents reproduced a known RSA-1024 forgery with 3.5x less CPU time than earlier academic work.</cite> <cite index="19-5">In In Plain English's account, the rerun took 391 core-years where the earlier work logged 1,380, which it treats as an implementation gain rather than a new attack.</cite>
The company says systems built on 2048-bit keys or larger fall outside these results, and the experiments were confined to an isolated setting that does not interact with running infrastructure. <cite index="18-10">The DeepMind papers are early working versions and numbers may change after review.</cite> <cite index="14-11">The work was led by DeepMind researchers Dj Dvijotham and Abhradeep Thakurta.</cite> <cite index="13-2">The agents were guided by cryptanalysts Nadia Heninger and Emmanuel Thomé.</cite>
Why it matters
For teams building AI companies, the result is a concrete case of agentic systems contributing to cryptanalysis, a task that once required small specialist groups. DeepMind's own framing centers on speed, with agents compressing work that would normally take years into weeks. The findings do not show that RSA has fallen in general, but they do show that the cost of probing old key sizes is dropping. Products that still depend on 1024-bit RSA should plan a migration now, since these results add a documented reason to retire those keys. Security reviews should assume attackers can use comparable agents, which moves inventories of legacy keys in firmware, certificate pools and hardware security modules up the priority list. For investors, diligence on AI-driven security claims and on a target's migration readiness is now a more concrete question than before.
