Anthropic proves security properties in 123 open-source libraries with Claude

Claude: Claude is Anthropic's frontier large language model designed for advanced reasoning and code-related tasks. In this news, it is the primary AI system used to autonomously generate formal proofs of security properties for over one hundred real-world open-source libraries at the RISC-V assembly level. The work highlights how Claude can also identify and patch long-standing bugs during the proof generation process. Theorem: Theorem is a collaborator focused on formal verification tools and semantics for software analysis. In this news, it partnered with Anthropic to develop and apply a pipeline that translates RISC-V binaries and Linux process models into mathematical specifications for Claude to prove. Theorem is also publishing a Claude skill and offering verification services for other open-source projects. Luna Tong: Luna Tong is a researcher contributing to AI-assisted formal verification efforts. In this news, she is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Ava Howell: Ava Howell is a researcher contributing to AI-assisted formal verification efforts. In this news, she is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Chris Rohlf: Chris Rohlf is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Jason Gross: Jason Gross is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Keane Lucas: Keane Lucas is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Sami Jawhar: Sami Jawhar is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Aengus Lynch: Aengus Lynch is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Luke Champine: Luke Champine is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Dino Distefano: Dino Distefano is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Nicholas Carlini: Nicholas Carlini is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Zygi Straznickas: Zygi Straznickas is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Evyatar Ben Asher: Evyatar Ben Asher is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Harshikaa Agrawal: Harshikaa Agrawal is a researcher contributing to AI-assisted formal verification efforts. In this news, she is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Rajashree Agrawal: Rajashree Agrawal is a researcher contributing to AI-assisted formal verification efforts. In this news, she is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Nickolai Zeldovich: Nickolai Zeldovich is a researcher contributing to AI-assisted formal verification efforts. In this news, he is listed as a co-author of the study demonstrating large-scale use of Claude to prove security properties in production open-source software. Proof Reusability: Shared formal models and proof components allow verification efforts across different programs to build on each other, reducing redundant work as more codebases are formalized. Bug Discovery Benefit: The process of attempting formal proofs with AI can uncover latent bugs in mature, heavily tested libraries that traditional testing methods have missed. AI Verification Scaling: Frontier AI models are enabling the generation of formal proofs for real-world software at assembly level across many open-source projects in weeks rather than years of expert effort.