Security
Google DeepMind says human-guided AI agents found two weaknesses in legacy 1024-bit RSA within weeks; it says 2048-bit keys are not affected, and the results come from working papers
DeepMind says human-guided AI agents found two weaknesses in obsolete RSA-1024; RSA-2048 is unaffected. No independent review seen yet.
On November 6, 2026, Google DeepMind said that AI agents working under human supervision found two weaknesses in legacy 1024-bit RSA: one in signature padding and one in factoring many keys at once. The company says RSA-2048 and modern schemes are not affected. We read the announcement through the post by DeepMind's Pushmeet Kohli and press relays, not the working papers themselves, and found no assessment by independent cryptographers.
The essentials
- The announcement: DeepMind says agents guided by humans made two cryptographic advances in weeks, working with cryptographers including Nadia Heninger and Emmanuel Thomé.
- Signatures: the agents found an algorithm that is about 1,600 times better than the previous best attack in theory, in a setting where the attacker controls a small part of the signed data. They also reproduced a recent attack on an already-solved key using about 3.5 times less CPU time.
- Batch factoring: a new method cuts the average cost of each additional 1024-bit key by about 30 times. Per WebProNews it is a paper result and not yet implemented.
- Scope: DeepMind says legacy 1024-bit RSA only. Keys of 2048 bits and modern schemes are not affected.
What is claimed and what is not shown
The 1,600x figure is a theoretical gain in one setting, and the 3.5x figure is a measured saving on a key that academics had already attacked. The 30x figure is an estimate for bulk factoring. None of this means a real-world key has been broken, and we saw no report of one.
How to read the weeks versus years framing
DeepMind and Kohli contrast weeks with the years or decades such advances have taken. That comparison is the company's own. The work was human-guided: WebProNews describes about 20 short messages from a researcher in one run. Two working papers are cited, and we saw no peer review.
What is not established
- Assessment by independent cryptographers: not found.
- Peer review of the two working papers: not seen.
- Any real 1024-bit key broken with these methods: none reported.
- Real-world cost of the batch-factoring method: not implemented.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Announcement | Google DeepMind, November 6, 2026, via Pushmeet Kohli's post | X (Pushmeet Kohli) |
| Team | Dj Dvijotham, Abhradeep Thakurta, with Nadia Heninger and Emmanuel Thomé | WebProNews |
| Signature result | About 1,600x in theory in one setting; 3.5x less CPU time on a solved key | Interesting Engineering |
| Batch factoring | About 30x cheaper per additional key, on paper | Interesting Engineering |
| Scope | Legacy 1024-bit RSA; DeepMind says 2048-bit is not affected | FourWeekMBA |
FAQ
Is RSA-2048 broken?
No. DeepMind says its results do not affect RSA-2048 or modern schemes.
Did AI find this alone?
No. DeepMind describes agents guided by human supervision and outside cryptographers.
Has a real 1024-bit key been broken?
We saw no such report. The signature result was reproduced on an already-solved key.
Method and limits
Sources checked on November 7, 2026. We did not read the working papers and relied on Kohli's post and press relays; one relay page could only be seen in search excerpts. No independent expert assessment was found. This is not security advice.
Sources
- X (Pushmeet Kohli): Pushmeet Kohli on X: two cryptographic advances in legacy 1024-bit RSA (verified 2026-11-07)
- FourWeekMBA: DeepMind AI Agents Find 2 Weaknesses in Legacy 1024-Bit RSA (verified 2026-11-07)
- Interesting Engineering: Google DeepMind finds 1,600x faster way to expose encryption flaws (verified 2026-11-07)
- WebProNews: Google DeepMind's AI Agents Find Two Weak Spots in Legacy 1024-Bit RSA (verified 2026-11-07)
- Digg: Google DeepMind announces AI-assisted research into two weaknesses in legacy 1024-bit RSA (verified 2026-11-07)
Spotted an error in this article? Report it · Our corrections policy