Claude Verifies 123 Libraries with Formal Proofs

In this article you will see:

• What it means to formally verify a program.
• The scale of the study conducted by Anthropic.
• The limits of the proofs described in the report.

How Claude and Rocq produce formal proofs

Formal verification is the use of mathematical proofs to demonstrate specified properties of a program. Anthropic combined Claude with Rocq, a system that checks whether each proof follows the defined logical rules.

The team describes the work in its engineering report on formal verification. Claude produces the proofs, and the Rocq kernel verifies the steps against the specification.

Image: Anthropic.

Anthropic applied the method to 123 libraries

The report covers 123 open source libraries and 239 test entry points from OSS-Fuzz. The team says the proofs in Rocq total about 180 million lines, not lines of original source code.

The binaries were compiled for RISC-V. The proof of absence of memory faults and unexpected terminations holds for inputs up to 64 MB and depends on specific assumptions about the operating system, memory and execution.

Fuzzing is a technique that tests programs with varied inputs to find bugs. For another example of AI applied to code security, see how OpenAI described the discovery of critical vulnerabilities in PHP](https://bridgecode.app/blog/newsletter/gpt-5-6-sol-ajuda-a-detectar-falha-critica-no-php).

What Claude's proofs do not guarantee

A proof covers the formalized properties and the recorded assumptions. It does not demonstrate that the entire system is free of any kind of failure.

Anthropic says it published reliable components of the project, complete proofs for some tests and specifications for other cases. The article does not announce a public tool that automates this entire flow without review.