DeepMind uses AI to find two weaknesses in 1024-bit RSA

$GOOGL  $ANTHROPIC  $ZEC #AI  #谷歌  #X 
On November 6, Google DeepMind announced two cryptanalysis studies in which AI agents, under human supervision, identified two weaknesses in older 1024-bit RSA encryption. DeepMind Senior Vice President of Science Pushmeet Kohli explained in a lengthy post on X that this shows such systems can crack it with less computing power than previously thought.
Kohli emphasized that the findings affect only the obsolete 1024-bit RSA. Encryption protocols that now almost universally use keys of 2048 bits or more remain secure.
DeepMind: AI makes in weeks advances that once took years or decades
RSA’s security rests on a mathematical “hard problem”: multiplying two large prime numbers is easy, but reversing the process and factoring the product back into primes is extremely difficult. The cryptography community uses the retired RSA-1024 as a testbed to continually evaluate this type of assumption.
Kohli said breakthroughs of this kind used to take years or even decades, allowing defenders to assume that the security margin changed slowly. But AI agents made two important advances in just a few weeks.
The research was led by DeepMind’s Dj Dvijotham and Abhradeep Thakurta, with collaborators including Nadia Heninger of the University of California, San Diego, and the French National Institute for Research in Digital Science and Technology (Inria), as well as Inria’s Emmanuel Thomé.
Weakness 1: If an attacker controls half the content, signature padding is useless
The first finding concerns digital signatures—the mechanism computers use to verify that software updates, documents, or websites really come from the claimed source. Many signature standards add preset data, known as “padding,” before signing to prevent forgery.
DeepMind found that the key factor is how much of the signed block an attacker can control. Once they control around half or more, padding offers no protection at all, and forging a signature is just as easy as it would be without padding.
The team tested its method in an isolated simulation environment and did not touch any live systems. The AI agent repeated an experiment using a 1024-bit key that had already been broken in previous academic research, using 3.5 times less CPU time and less than one-thousandth of the computational effort required to break the key itself.
Forgery is harder when attackers can control less content, such as just a 32-byte digital fingerprint. Even so, the team found a new algorithm that, according to theoretical estimates, is about 1,600 times faster against 1024-bit keys than the previous best method.
Weakness 2: Break many keys at once, reducing the cost per key by about 30 times
The second finding involves factoring many numbers of the same length at once. RSA keys typically use standard lengths, such as 1024 or 2048 bits. An attacker seeking to break many keys can spread the cost of one large initial computation across all of them.
The AI agent found a new algorithm that, compared with previous batch methods, reduces the average cost of factoring each additional key in a large batch of 1024-bit keys by about 30 times. Technical papers on both findings have been published on the DeepMind website (digital signatures, key factoring).
“We can no longer treat cryptographic difficulty as a fixed assumption”
Kohli wrote that modern cryptographic systems, from encrypted communications to digital currencies, rely on computational problems of this kind. “In a new era of AI-driven mathematics and computing, we can no longer treat cryptographic difficulty as a fixed, unchanging assumption.” DeepMind said it published the research to help the cybersecurity community assess cryptographic strength more accurately and strengthen standards before threats emerge.
Examples of AI finding cryptographic vulnerabilities have already appeared in the crypto industry. This June, an Anthropic model uncovered a vulnerability in Zcash’s privacy pool that had been hidden for four years, in just a few days. See the earlier report.
This article, “DeepMind Uses AI to Find Two Weaknesses in 1024-bit RSA,” first appeared on .
【Further Analysis】
【My Further Perspective】
The core takeaway from this study is the “accelerating effect of AI on solving difficult mathematical problems.” Cryptographic security assumptions have traditionally been based on the relatively slow evolution of algorithms and computing power. But AI agents can compress breakthroughs that once took years or even decades into just a few weeks. This means the “buffer time” for security defenses has been drastically reduced. The urgency of adopting higher-level encryption standards—such as post-quantum cryptography (PQC) or longer key lengths—will increase significantly, whether for Web2 internet infrastructure or Web3 cryptographic protocols.
【Market Impact】
1. Cryptocurrency market: Overall impact is 【neutral】. Major coins such as Bitcoin (BTC) and Ethereum (ETH) use elliptic-curve cryptography (Secp256k1), not the RSA system discussed in this study, so their underlying cryptographic security has not been directly affected. However, privacy coins such as Zcash (ZEC), mentioned in the article, and early blockchains with older technical architectures could face slightly 【bearish】 market sentiment in the short term. Funds may favor projects that proactively upgrade to newer zero-knowledge proofs or higher security standards.
2. U.S. stock market: 【Bullish】 for Alphabet (GOOGL) and the cybersecurity sector (such as Palo Alto Networks and CrowdStrike). DeepMind’s findings demonstrate Google’s depth in AI applications. At the same time, the potential risk of AI-enabled cryptographic attacks could prompt companies around the world to accelerate cybersecurity spending.
3. Taiwan stock market: 【Bullish】 for stocks related to information security and systems integration (such as Zero One Technology and Systex). Demand for security system upgrades and cryptographic transitions could drive additional business in the future.
【What to Watch Next】
If future research on AI agents extends to 2048-bit RSA or even elliptic-curve algorithms, it could trigger broader industry alarm. Keep monitoring the pace at which the NIST post-quantum cryptography standards are adopted worldwide, as well as the security upgrade roadmaps of major blockchain communities.
【Risk Reminder】
Investors should beware of unnecessary market panic (FUD) triggered by claims that “AI can break encryption.” The RSA-1024 system affected in this study had already been retired from mainstream applications. Modern mainstream financial and internet protocols remain secure, so do not overreact and sell indiscriminately.