Google DeepMind uses AI to find RSA encryption flaws 1,600x faster

Accelerating cryptanalysis

Google DeepMind researchers have developed a method to expose weaknesses in legacy RSA encryption at a rate 1,600 times faster than previous techniques. This work was announced by Pushmeet Kohli, the company's Vice President and Chief Scientist, following a study involving AI agents. The project highlights how artificial intelligence can be used to navigate complex mathematical problems that have historically required years of human effort.

Cryptanalysis is the specific science of testing encryption systems to find vulnerabilities before malicious actors can exploit them. For many decades, the field has moved slowly because the mathematical hurdles were so significant that experts could only make incremental progress. By using AI agents to explore these problems, the researchers were able to compress the timeline of discovery significantly while working under human supervision.

The study, which involved collaborators from the University of California San Diego and France’s Inria, focused specifically on the 1024-bit RSA standard. This is an older encryption method that has largely been replaced by more robust 2048-bit keys. While the results are significant for the field of mathematics, the researchers noted that modern encryption used for banking and messaging remains secure.

Weaknesses in digital signatures

One of the primary findings involves the way digital signatures verify the legitimacy of software updates and electronic documents. These schemes often use a technique called padding, which adds preset information to data before it is signed. While padding was intended to prevent forgery, the researchers found that its effectiveness is not as strong as previously thought when an attacker can influence parts of the message.

The AI agents discovered that if an attacker can control roughly half of the signed data block or more, the padding provides virtually no extra protection. In these specific scenarios, forging a digital signature becomes mathematically no more difficult than attacking a system that has no padding at all. This finding helps security experts understand the limits of current signature protections in specific environments.

To verify these findings, the team tested the AI-generated method in a sandbox environment using an existing RSA-1024 benchmark. The AI-driven implementation reproduced the expected results while using 3.5 times less CPU time than previous methods. Furthermore, it required a fraction of the processing power needed to factor the benchmark key itself, proving the efficiency of the new approach.

Batch factoring efficiency

The second major breakthrough addresses the batch factorization of multiple RSA keys that share the same bit length. Because many systems use standardized key sizes, it is theoretically possible for an attacker to share the cost of heavy computations across many different keys at once. The AI-discovered algorithm improves the asymptotic efficiency of this process, making it much cheaper to target multiple keys simultaneously.

Specifically, the new method reduces the average price in computing power required to break each extra 1024-bit RSA key by about 30 times compared to previous batch techniques. This means that as the number of keys increases, the cost per key drops much faster than it did with older methods. This discovery highlights how AI can optimize the way researchers and attackers alike approach large-scale mathematical problems.

The research underscores a growing trend where AI becomes increasingly capable of tackling advanced math that was once thought to be computationally prohibitive. By identifying these shortcuts, the study provides a clearer picture of how quickly the mathematical assumptions that underpin digital security can be evaluated. This allows for a more proactive approach to identifying flaws in the foundations of our digital world.

The future of security

Current encryption standards are still safe, but the team suggests that AI-driven cryptanalysis could become a vital way to test future systems before they face real-world threats. For decades, the field moved slowly, giving experts time to update standards as computers got faster. AI is now making that process happen much more quickly.

The study, led by Dj Dvijotham and Abhradeep Thakurta, produced two major findings involving legacy RSA-1024 cryptosystems. The AI agents even found a way to improve on the best known attacks by roughly 1,600 times for 1024-bit RSA keys, even when the attacker only manages to influence a small piece of the data, like a 32-byte fingerprint.

Kohli noted that these results show why it is important to keep checking the math behind our security as AI gets better at solving hard problems. This shift means that security experts must now account for AI as a primary driver of cryptanalysis. By using these tools now, the industry can stay ahead of potential threats by identifying weaknesses before they can be weaponized by malicious actors.