Developing New Methods for Breaking Legacy Encryption
Google DeepMind has announced new research findings demonstrating the ability of AI agents, under human supervision, to reduce the computational cost required to attack legacy RSA-1024 encryption systems. This research was conducted in collaboration with cryptography experts from the University of California, San Diego, and the French National Institute for Research in Digital Science and Technology (Inria).
Pushmeet Kohli, Vice President at Google DeepMind, emphasized that these results do not pose a threat to modern encryption protocols that rely almost universally on 2048-bit keys or larger; rather, the implications are limited to legacy systems that still utilize the 1024-bit standard.
Research Findings: Digital Signatures and Key Factorization
The research revealed two key findings:
- Digital Signature Forgery: Researchers found that the effectiveness of "padding" in protecting digital signatures depends directly on the proportion of the block controlled by the attacker. When an attacker controls half the block or more, the padding loses its protective capability entirely, which applies to the PKCS#1 v1.5 format used in Hardware Security Modules (HSMs). In an experimental test, the AI succeeded in reducing the computation time required to forge a signature by 3.5 times compared to previous studies.
- Key Factorization: The agents developed an algorithm that makes the process of factoring RSA-1024 keys approximately 30 times cheaper on average when dealing with large sets of keys, by sharing the initial computational cost among them.
Why Does This News Matter?
The importance of these results lies in the fact that they challenge the notion of "cryptographic hardness" as a fixed assumption. This development illustrates how AI can re-evaluate the mathematical assumptions upon which cybersecurity standards have been built for decades. The researchers point out that using AI to rigorously test these assumptions is a necessary step to ensure that modern encryption systems remain secure in the face of evolving computational capabilities.
In practical terms, these attacks do not lead to the direct recovery of the private key, and they require a massive number of pre-signature queries, which limits their applicability in real-world scenarios against systems that hash messages internally.