Anthropic says its Claude models have verified security properties across 123 open-source libraries in a matter of weeks. The work was carried out by Anthropic’s Frontier Red Team with Theorem Labs and focused on assembly-level verification, which examines the low-level instructions actually executed by a processor. Traditional formal verification had required years of specialist effort, making mathematical proofs impractical for many real-world projects. An earlier analysis of 25 open-source projects, covering more than 5 million lines of code, identified 493 issues, including 94 cases involving disabled security controls. The researchers argue that demonstrating code is free of specified bugs could strengthen defensive security, although the method’s usefulness will depend on reproducibility and on the completeness of the specifications being verified.
Source: Read the original article
