Researchers at Google DeepMind have used AI agents to discover cryptographic techniques that uncover vulnerabilities in legacy RSA encryption up to 1,600 times faster than previous methods.
Pushmeet Kohli, Vice President and Chief Scientist of DeepMind, announced this breakthrough in a social media post today.
The research represents a significant advance in the field of cryptanalysis while leaving modern encryption standards untouched.
No threat to modern encryption
The research focused on 1024-bit RSA, an outdated encryption standard that has largely been replaced by stronger 2048-bit or larger keys. According to Google DeepMind, the findings do not pose a threat to modern encryption used to protect online banking, messaging apps, software updates or digital currencies. Instead, they show how AI can significantly speed up the assessment of the mathematical assumptions underlying digital security.
Cryptanalysis is the science of testing encryption systems by looking for vulnerabilities before malicious actors can exploit them. For decades, advances in the field typically took years or even decades, allowing security experts to gradually strengthen encryption standards as computing capabilities evolved.
The company said its AI agents, working under human supervision, shortened that time frame by quickly examining math problems that traditionally required extensive manual research.
The study, led by Dj Dvijotham and Abhradeep Thakurta in collaboration with researchers from the University of California San Diego and France’s Inria, produced two important findings about older RSA-1024 cryptosystems.
“Padding” prevents counterfeits
The first concerns digital signatures, which allow computers to verify that software updates, websites and electronic documents come from legitimate sources.
Digital signature schemes typically add additional preset information called “padding” before signing data. Padding has long been thought to prevent forgery, but researchers have not fully quantified its effectiveness when attackers can influence parts of the signed message.
During the investigation, the AI agents found that padding provides virtually no additional protection once an attacker controls approximately half or more of the signed data block. Under these conditions, forging a digital signature is mathematically no more difficult than attacking an unbuffered system.
To validate the discovery, the researchers tested the method in a sandbox environment using a previously solved RSA-1024 benchmark from previous academic work. The AI-generated implementation reproduced the result while using 3.5 times less CPU time and less than a thousandth of the computing power required to factorize the benchmark key itself.
Even if attackers only control a small portion of the signed data, such as a 32-byte digital fingerprint, the AI agents identified a new algorithm that theoretically improves the previous best attack for 1024-bit RSA keys by about 1,600x.
AI reduces the computational effort of cryptanalysis
The second breakthrough aims to batch factorize multiple RSA keys with the same bit length. Because many RSA deployments use standardized key sizes, attackers could theoretically spread the cost of expensive computations across numerous keys.
The AI-discovered algorithm improves the asymptotic efficiency of this process, reducing the average computational cost of factoring each additional 1024-bit RSA key by approximately 30x compared to previous batch techniques.
Kohli said the results highlight the need to continually re-evaluate cryptographic assumptions as AI becomes increasingly capable of solving advanced mathematical problems.
While today’s encryption standards remain secure, the researchers argue that AI-powered cryptanalysis could become an essential tool for stress testing future cryptographic systems before emerging threats compromise them.
