Researchers from Google DeepMind, in collaboration with cryptographers from UC San Diego and Inria, have demonstrated that 1024-bit RSA encryption is more vulnerable to targeted attacks than previously understood. By utilizing AI agents to optimize computational strategies, the team identified methods to reduce the cost of forging digital signatures and factoring RSA-1024 keys. Google has emphasized that these findings are limited to the obsolete 1024-bit standard and do not impact modern 2048-bit or larger encryption protocols, which remain secure.
The first major finding concerns the forgery of digital signatures, specifically those utilizing the PKCS#1 v1.5 padding format common in hardware security modules and cryptographic signing interfaces. The research indicates that the effectiveness of padding as a security measure is directly tied to the amount of data an attacker can control within a signed block. When an attacker influences roughly half or more of the block, padding provides no meaningful protection against forgery.
In a controlled, sandboxed environment, the AI agents successfully replicated a prior academic forgery attack on a 1024-bit key. By designing, executing, and debugging the computation with minimal human intervention—requiring only about 20 approvals over a four-day period—the agents reduced the necessary CPU time to 391 core-years. This represents a 3.5-fold improvement over the 1,380 core-years required in the original study.
Despite these efficiency gains, the researchers confirmed that the private key remained uncompromised during the forgery process. Furthermore, the attack is constrained by specific requirements: it necessitates approximately 4 billion signing queries to be performed in advance and is ineffective against devices that perform internal message hashing. For scenarios where the attacker controls a smaller portion of the block, such as a 32-byte hash, the team developed a new algorithm that is theoretically estimated to be 1,600 times more efficient than previous methods.
The second finding addresses the broader challenge of factoring multiple RSA-1024 keys. The AI agents developed an algorithm that allows an attacker to amortize the cost of a large, initial computation across multiple keys. This approach is estimated to make the factoring of each subsequent 1024-bit key roughly 30 times cheaper than the best-known previous techniques. However, this remains a theoretical estimate, as the method has not yet been implemented and assumes a significant upfront computational investment.
The researchers explicitly noted that factoring a single 1024-bit key remains a massive undertaking, requiring an estimated 500,000 to 1 million core-years of computation. The study highlights a shift in the landscape of cryptographic security, suggesting that as AI-driven mathematical and computational capabilities advance, long-held assumptions regarding the difficulty of certain cryptographic tasks must be rigorously re-evaluated.
Google officials stated that the project serves as a proactive measure to ensure the continued integrity of modern encryption standards. By using AI to stress-test existing cryptographic assumptions, the research aims to identify potential weaknesses before they can be exploited in real-world scenarios. The findings underscore the importance of transitioning away from legacy 1024-bit RSA systems, which are increasingly susceptible to optimization-based attacks.
