Google DeepMind says AI agents found two weaknesses in legacy 1024-bit RSA
This is DeepMind's own announcement as reported by one outlet; this edition did not read DeepMind's post, which was not found at the address tried, or the underlying papers. The 1,600-fold figure is a theoretical estimate, not a measured attack. DeepMind says the findings do not threaten 2048-bit or larger keys used for banking, messaging and software updates. Researchers guided the agents, so this is not autonomous discovery, and the share of the work done by the agents is not stated in the source read.
What was claimed?
Google DeepMind announced on November 6, 2026 that AI agents, working under human supervision with researchers from UC San Diego and France's Inria, produced two findings on legacy 1024-bit RSA encryption, Interesting Engineering reported. First, the agents found that once an attacker controls roughly half of a signed data block, the padding added before signing offers almost no extra protection against forgery; a test on an already-solved RSA-1024 benchmark reproduced the earlier result with 3.5 times less CPU time. Where an attacker controls only a small part, such as a 32-byte fingerprint, DeepMind says a new algorithm improves on the previous best attack by about 1,600 times in theory. Second, it says a new method cuts the average cost of factoring each additional 1024-bit key in a batch by about 30 times compared with earlier batch techniques.
What does the evidence support?
Company claim, independent confirmation unresolved
Advances in breaking encryption have usually taken years, and DeepMind says AI agents compressed some of that work into weeks. If such tools speed up attacks, older systems still using 1024-bit keys become riskier, and standards bodies may need to re-test assumptions more often. The same tools can help defenders stress-test new systems before attackers do.
What are the limits?
This is DeepMind's own announcement as reported by one outlet; this edition did not read DeepMind's post, which was not found at the address tried, or the underlying papers. The 1,600-fold figure is a theoretical estimate, not a measured attack. DeepMind says the findings do not threaten 2048-bit or larger keys used for banking, messaging and software updates. Researchers guided the agents, so this is not autonomous discovery, and the share of the work done by the agents is not stated in the source read.
Claim or evidence date: 2026-11-06. Last verified: .
Interesting Engineering · Google DeepMind agents expose weaknesses in older encryption methods ↗
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →