Aviz is a multidisciplinary team of Inria and Université Paris-Saclay aiming at improving the analysis and visualization of large and complex datasets by combining analysis methods with interactive visualizations. Learn more...

Researcher homepage MOEX INRIA

Artificial intelligence (AI) agents developed by Google DeepMind have shown new ways to make attacks against legacy RSA-1024 encryption systems significantly cheaper.  The research was conducted by Google DeepMind in collaboration with cryptography researchers from the University of California San Diego and Inria.  RSA is a widely used public-key cryptographic algorithm that relies on the mathematical difficulty of factoring large numbers. Although RSA-1024 was once considered sufficiently secure, advances in computing and cryptanalysis have made it unsuitable for protecting sensitive information. Modern systems generally use RSA keys of 2048 bits or longer. The first major finding focused on digital signature forgery involving the PKCS#1 v1.5 padding format. Padding is additional data inserted into a message before signing to ensure that it meets the required block size. Researchers examined how this protection behaves when an attacker can influence the content of the message being signed. The team found that when an attacker controls approximately half or more of the signed data block, the padding may no longer provide additional protection against forgery. This weakness is relevant to the PKCS#11 signing interface used by cryptographic devices, including hardware security modules. To show the improvement, DeepMind's AI agents reproduced an earlier academic signature-forgery attack against an RSA-1024 key. The agents reduced the required computing time from approximately 1,380 CPU core-years to 391 CPU core-years, making the process around 3.5 times more efficient. Two AI agents reportedly designed, executed and debugged the computation in less than four days, with approximately 20 messages from a human researcher, most of which involved approving their actions. Researchers also developed a more efficient approach for situations where attackers control a smaller portion of the signed message, such as a 32-byte hash. Their theoretical calculations suggest that the proposed method could be approximately 1,600 times more efficient than the previous best approach for certain RSA-1024 signature-forgery scenarios. The second major finding concerns RSA key factoring. RSA security depends on the difficulty of breaking down a large number into its original prime factors. The researchers developed an algorithm that could make factoring each additional RSA-1024 key approximately 30 times cheaper on average after an expensive initial computation. However, this technique remains theoretical and has not yet been implemented. The initial computation would also require substantial resources, meaning the findings do not show an immediate practical attack against deployed encryption systems. The researchers also highlighted important limitations of the signature-forgery technique. It requires approximately four billion signing queries in advance, does not recover the private key and does not work against devices that internally hash messages before signing. According to Pushmeet Kohli, Vice President at Google DeepMind and Chief Scientist at Google Cloud, modern cryptographic protocols generally use RSA keys of 2048 bits or larger and remain secure against the demonstrated techniques.

Notre modèle

Google DeepMindは、人間の監督下で動作するAIエージェントが、1024ビットRSA暗号システムへの攻撃コストを下げる2つの手法を発見したと発表しました。 カリフォルニア大学サンディエゴ校およびInriaの暗号研究者と共同で実施した今回の研究により、旧式のRSA-1024に対する攻撃は、従来の想定より少な

Quantum Zeitgeist

C’est une prestigieuse mise à l’honneur pour notre collègue Christophe Botella, chercheur à l’Inria et membre de l’équipe Pl@ntNet. En effet, il sera reçu au Collège de France le 26 novembre prochain ...

A New Warning for Legacy Encryption as AI Accelerates Cryptographic Research Google DeepMind has revealed research demonstrating how artificial intelligence agents, working under human supervision, can discover and optimize mathematical techniques that reduce the computational cost of attacking legacy RSA-1024 cryptographic systems. The findings, developed alongside cryptographers from the University of California San Diego and France's Inria research institute, suggest that certain attacks against 1024-bit RSA encryption may be considerably more efficient than previously understood.

El equipo de Ciencia de Google DeepMind trabajó con criptoanalistas de la Universidad de California San Diego y del Inria francés, con Dj Dvijotham y Abhradeep Thakurta a la cabeza, y dice haber llegado a los dos resultados en semanas. El primero es sobre padding, el bloque de datos preestablecido…

a-Shellはフランス国立情報学自動制御研究所(INRIA)の研究員で開発者のNicolas Holzschuchさんが2019年から公開しているターミナルエミュレーターで、iOSのSandbox内で安全に動作し、Unixコマンドの学習やPythonを利用したデータ処理、Lua, JavaScript, C/C++言語のプログラミングが出来ますが、このa-Shellが。