Anthropic's Claude now writes code and proves it safe automatically

Anthropic has demonstrated that its Claude model can generate mathematical proofs verifying that existing code behaves correctly. The effort covered 123 open-source libraries and produced approximately 180 million lines of proof code, applying formal verification—a rigorous mathematical method—to real-world software at scale.

The project is significant because formal verification has traditionally been prohibitively expensive and labor-intensive, requiring specialized human mathematicians. By automating proof generation with an LLM, Anthropic is testing whether large language models can handle tasks that demand precise logical reasoning rather than just pattern matching.

During the process, Claude identified bugs that had survived years of conventional testing, suggesting that LLM-driven verification could complement existing quality-assurance practices. This points toward a potential new application for frontier models: not just writing code, but certifying its correctness.

If this approach scales reliably, it could lower the barrier to formal verification for open-source and commercial software, improving long-term reliability. However, the practical adoption of LLM-generated proofs will depend on whether the outputs are consistently sound and whether developers trust automated verification over manual review.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.