Anthropic and its partner Theorem have utilized the Claude model to perform formal verification of existing software. The system translates crash tests into mathematical problems, analyzes compiled machine code, and generates proofs that are subsequently validated by the Rocq program.
During the project, Claude processed test programs from 123 open-source libraries within Google's OSS-Fuzz project, generating approximately 180 million lines of proof code in a few weeks. The model identified and corrected hundreds of bugs, including a logical error that had persisted for over a decade.
The verification process is currently limited to specific conditions, including single-threaded programs, particular processor architectures, and inputs under 64 MB. Despite these constraints, the team discovered as many security bugs in a PDF reader in one week as are typically reported in a year.
While the compute cost remains high, with individual proofs requiring hundreds of millions of tokens, efficiency is improving. The Opus 5.5 model demonstrated the ability to solve half of the proofs using 40 percent fewer tokens than its predecessor.
Anthropic aims to make AI-driven formal verification a practical alternative to standard testing. Future plans from Theorem include applying this methodology to critical infrastructure software, such as OpenSSL, glibc, and cURL.
