Skip to content
Live
Buying GuidesBest Laptops to Buy Right Now: MacBook Air M5, MacBook Neo, XPS 13 and OmniBook 5 by BudgetFrontiersGoogle Books Blue Origin's New Glenn to Launch Rack-Scale AI Satellites in Early 2029LifestyleIntuit Sells Mailchimp to Bending Spoons for $3.68 Billion, Under a Third of What It PaidMacGemini for Mac Can Now Click, Type and Fill Forms for You in a Background WindowiPhoneApple Is Building an AI Agent That Can Control Apps and Write Code, Job Listings SuggestApple CornerUK Tribunal Forces the Home Office to Admit It Served Apple a Secret Encryption NoticeLaptopsLenovo Posts a Record $30.2 Billion Quarter but Slips Into an Operating Loss on a Warrant ChargeAndroidGoogle Will Require Foldable-Ready Android Apps on Google Play From April 2027Artificial IntelligenceSignal Blocks Screenshots by Default to Keep Siri AI and Gemini From Reading Your ChatsGamingGTA 6 Takes $3.79 Billion in Day-One Sales, Take-Two Says, the Biggest Entertainment Launch EverSmartphonesHuawei Launches OSTA, a New Phone Brand, With the Wide-Screen O1 Pro From 5,499 YuanComputingNvidia Plunges 14% in a Day as an SEC Probe Turns to Its AccountingBuying GuidesBest Laptops to Buy Right Now: MacBook Air M5, MacBook Neo, XPS 13 and OmniBook 5 by BudgetFrontiersGoogle Books Blue Origin's New Glenn to Launch Rack-Scale AI Satellites in Early 2029LifestyleIntuit Sells Mailchimp to Bending Spoons for $3.68 Billion, Under a Third of What It PaidMacGemini for Mac Can Now Click, Type and Fill Forms for You in a Background WindowiPhoneApple Is Building an AI Agent That Can Control Apps and Write Code, Job Listings SuggestApple CornerUK Tribunal Forces the Home Office to Admit It Served Apple a Secret Encryption NoticeLaptopsLenovo Posts a Record $30.2 Billion Quarter but Slips Into an Operating Loss on a Warrant ChargeAndroidGoogle Will Require Foldable-Ready Android Apps on Google Play From April 2027Artificial IntelligenceSignal Blocks Screenshots by Default to Keep Siri AI and Gemini From Reading Your ChatsGamingGTA 6 Takes $3.79 Billion in Day-One Sales, Take-Two Says, the Biggest Entertainment Launch EverSmartphonesHuawei Launches OSTA, a New Phone Brand, With the Wide-Screen O1 Pro From 5,499 YuanComputingNvidia Plunges 14% in a Day as an SEC Probe Turns to Its Accounting
Artificial Intelligence

The reported gains apply to older 1024-bit keys only. Google's science chief says RSA 2048 is not affected.

P

By

· 3 min read

Google DeepMind Says AI Agents Found Two Weaknesses in Legacy 1024-Bit RSA, but Modern Encryption Is Unaffected

Google DeepMind says AI agents have found two significant weaknesses in a legacy form of RSA encryption, the first time the company has publicly credited its agents with improving attacks on a widely deployed cryptographic scheme. The announcement came on Friday 6 November from DeepMind and Pushmeet Kohli, its vice president of science. The affected keys are 1024 bits long, a size the industry has been retiring for more than a decade.

What DeepMind announced

According to press coverage of the announcement, AI agents guided by human cryptanalysis experts found the two weaknesses within weeks. The first concerns signature forgery. Interesting Engineering and Byteiota both report that a new algorithm for 1024-bit keys is about 1,600 times more efficient than the previous best attack. The gain holds even when an attacker controls only a small part of the signed block, such as a 32-byte hash.

Interesting Engineering adds that when an attacker controls roughly half of the signed block or more, the padding used in the scheme gives virtually no extra protection. It also reports that in a sandbox, an AI-generated implementation reproduced an already solved academic RSA-1024 benchmark using 3.5 times less processor time. That detail appears in one outlet only.

The second weakness is about batch factoring. Both outlets report that the marginal cost of factoring each additional 1024-bit key in a large batch falls by about 30 times compared with earlier batch methods. In practice that matters to an attacker who holds many keys at once rather than one.

What is not affected

Kohli was explicit about scope. Coverage quotes him saying the results "do not affect modern encryption schemes such as RSA 2048". RSA keys of 2048 bits and above, which protect most current web traffic and software signing, are outside the findings.

The affected keys are old. Byteiota recalls that RSA-1024 was barred for new deployments in 2013, and that certificate authorities stopped issuing 1024-bit certificates in 2014. Systems still relying on them are mostly legacy devices, embedded hardware and long-lived internal infrastructure.

How much weight the numbers carry

The headline figures are reported rather than independently verified. We could not locate a DeepMind blog post carrying the 1,600 times and 30 times figures. They appear in press write-ups of a longer account that Kohli referred to on X. Byteiota says no production certificate has been forged, and Interesting Engineering notes that several results are theoretical and that validation took place in a sandbox. Treat the numbers as claims that outside cryptographers still need to check.

The work also builds on earlier human research. A related academic paper on forging 1024-bit RSA signatures in nearly SNFS time, by researchers at UC San Diego and INRIA including Nadia Heninger and Emmanuel Thomé, was posted to the IACR preprint archive after being received on 20 September. It reports about 1,380 CPU core-years of computation over five months. Byteiota names Heninger and Thomé as the guiding human cryptanalysts for the DeepMind effort. That paper predates the 6 November announcement and should not be read as the announcement itself.

Why it matters beyond RSA

Cryptanalysis is slow, specialised work, and few people can do it well. If AI agents can speed up the search for weaknesses even in an old scheme, defenders and attackers both stand to gain from the same tools. Interesting Engineering names DeepMind researchers Dj Dvijotham and Abhradeep Thakurta in connection with the project, but the public record on how the agents were set up is thin so far.

The announcement lands amid a run of claims about AI and security. Gadgets365 reported earlier this week on OpenAI's statement that its own pipeline built 948 working Linux root exploits, a result that is also unverified outside the company. Taken together, they suggest AI labs are now treating offensive and defensive security research as a showcase for their agents.

What organisations should do

The practical advice is unchanged. Anyone still using 1024-bit RSA for signatures or key exchange should plan to move to 2048 bits or longer, or to elliptic curve schemes. Security teams in India that run older payment terminals, industrial controllers or legacy portals may want an inventory of where short keys remain. Nothing in the announcement says any live system has been broken.

What to watch next

Watch for a full technical write-up from DeepMind with the methods and test conditions, and for responses from independent cryptographers who can reproduce the efficiency gains. Standards bodies will also be worth following if the results prompt tighter guidance on padding schemes for signatures.

Published 8 November 2026. Gadgets365 will update this article as more information becomes available.

P

Editor

Pravin Kankane

Pravin Kankane is the editor of Gadgets365, an independent technology publication based in Pune, India. He covers smartphones, AI, Apple, Android and consumer tech launches in India.

Keep reading

Related stories