We’re sharing new research demonstrating how AI agents can accelerate the science of stress-testing encryption. Guided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems. Our VP Science @Pushmeet explains more ↓ xitter.twitterapi.xyz/pushmeet/status/211875…
Pushmeet Kohli (@pushmeet) profile picturePushmeet Kohli
@pushmeetNov 6
Advancing the frontier of cryptanalysis

Advancing the frontier of cryptanalysis

Today, we’re sharing new research from @GoogleDeepMind showing how AI agents can accelerate cryptanalysis, or the science of stress-testing encryption. Our team used AI agents, guided by human super…

Article · @pushmeet

Nov 6, 2026 · 5:20 PM UTC

316
208
158
1,570
174,416
RelevantRecentLikes
While modern 2048-bit systems remain safe from these specific exploits, this work highlights a key shift. As we enter a new era of AI-enabled math and computation, we can no longer rely on historical hardness assumptions and must continuously stress-test cryptographic standards.
25
7
4
131
40,841
Give a small cryptography team thousands of agents and suddenly all the weird branches nobody had time to chase become searchable. There are decades of old assumptions in legacy crypto worth running back through with that kind of search.
1
9
1,902
TL;DR 🔥 xitter.twitterapi.xyz/vishalsingh2972/status…
Vishal Singh (@vishalsingh2972) profile pictureVishal Singh
@vishalsingh2972Nov 6
This is actually crazy 🤯 Google used AI agents to find new ways to attack an older security system much faster and cheaper. 🔥 Why this matters: The security protecting our data relies on math that's extremely hard to crack. AI could help us discover weaknesses much earlier, before hackers can exploit them. Modern encryption is still safe from these attacks, but imagine how much AI could transform security in the coming years. Really excited to see how this changes cybersecurity.
Image from Vishal Singh (@vishalsingh2972)
1
5
672
agents stress testing old encryption is both useful and unsettling
1
282
We're glad to see AI helping the people who have to find the weak spots before an attacker does.
252
Such impressive research love seeing AI help make our digital world safer
127
“Too expensive to break” depends on the best attack we know. A better algorithm can shrink that safety margin without changing a single bit of the key. That’s what makes AI-assisted cryptanalysis so interesting. 💡
255
we need more of this to show the general public that AI is more than just a chatbot because that how most anti datacenter people view it right now
1
1
4
1,057
These don't seem tooooo bad yet but I can only imagine how annoying it will be in the future to keep old systems with built in legacy crypto secure given how fast the agents can iterate.
1
4
583
ok 1024-bit RSA is still out there? Didn't expect that.
1
286
the condition that matters for the signature result: the attacker needs a device that signs inputs of their choosing, about 4 billion of them in the 1024-bit run, all before the target is known. signers that hash the message inside the device are outside the model
1
1
371
the human in this sent like 20 messages total and most of them were basically ok go ahead
465
"Computationally hard" has always come with an asterisk (*) *using the best methods we know
338
good excuse to check your own keys today. ssh-keygen -lf ~/.ssh/id_rsa.pub shows the size, and if it says 1024 just make a new ed25519 one
2
397
1024-bit still running in production in 2026 is wild lol
302
Plain-language check for non-experts: the weak spot is old 1024-bit keys, and 2048-bit is still safe. So nothing to do today for ordinary people, and the work is on banks and websites to retire the old ones. Is that the right reading?
105
guided by humans is the part worth keeping. the muses still ask theirs before anything moves
263
Legacy RSA flaws aren't new info, researchers flagged risks for years now.
292
Curious how much of that was the model vs the humans.
2
417
@GoogleDeepMind @pushmeet
Image from Coinbase Predict (@CoinbasePredict)
1
5
209
1024-bit RSA fell, 2048-bit still holds.
7
谢谢您的研究
109
AI agents stress-testing legacy RSA with a human in the loop is the kind of research that makes me rethink what "safe enough" even means for old crypto.
47
Two addresses for the 'legacy' risk: device fleets, where each extra 1024-bit key gets about 30x cheaper on paper once the factory is paid for, and HSMs still exposing CKM_RSA_PKCS. Both sit in racks for decades. Worth pointing the agents at finding them.🙏
77
Are the two weaknesses fixable or is 1024 just done ?
51
Humans only sent approximately 20 messages, and all the code was AI-generated. The 2048-bit security was secure today, but this sense of security is changing.
16
Are there plans to stress-test recent PQC algorithms, too?
76
They’re coming for postquantum next…
5
70
Guided by humans is the key part here. I already run agents on almost everything
2
82
重点其实不是"AI 破了 RSA"，而是密码学审计从靠人类灵光一现，变成了可以批量搜索的假设检验。1024 位是历史包袱，真正该紧张的是那些还靠"没人有空算"活着的算法。
1
1
31
humans guiding it is the part i'd copy
1
76
2048-bit is safe from these ones, for now. But if agents keep finding cracks this fast, how long does that for now last?
1
24
The key words are guided by humans. Models that can search huge spaces of possibilities, with experts choosing where to look and checking the results, may be how a lot of science speeds up. Curious how much of the discovery will come from the model versus the guidance. Quantum might change the paradigm.
5
Turns out the best way to test a lock is to let the AI try picking it.
29
Interesting
31
broke 1024 bit rsa. password manager still asks twice
33
Load more replies