Anthropic says it partnered with Theorem to use Claude to formally prove useful security properties at the assembly code level across 123 open-source libraries

Code you use is now provably more secure against certain classes of bugs.

Anthropic has partnered with Theorem to leverage large language models (LLMs), specifically Claude, for formal software verification. This collaboration aims to automate the process of proving code security properties at the assembly code level, a task traditionally requiring extensive expert effort. The teams applied this method to 123 open-source libraries, successfully generating formal proofs for RISC-V assembly code. During this process, Claude identified and helped fix bugs that had previously evaded detection through trillions of test runs. The initiative demonstrates the potential of LLMs to scale formal verification, making it more feasible for real-world software development and enhancing security guarantees.

The partnership between Anthropic and Theorem signifies a major leap in automating software security. By using LLMs to generate formal proofs for assembly code, they are making rigorous verification accessible beyond specialized security applications. This move could fundamentally alter software development workflows, embedding stronger security guarantees directly into the development lifecycle.

This development impacts any user of open-source software, as the security of foundational libraries can now be assessed and proven at an unprecedented scale. The broader implication is a potential future where formal verification is a standard component of software release, significantly reducing vulnerabilities and increasing overall system robustness.