AI Model Certifies 123 Legacy Libraries After Weeks of Autonomous Proof Generation

Developers working with open-source code now have a new way to assess reliability that does not rely on manual code reviews. Anthropic announced on October 23 that its Claude model mathematically proved the correctness of 123 existing software libraries. The process required no new software writing; instead, the AI analyzed compiled machine code to construct extensive proofs that the code behaves as expected under specific conditions.

The verification effort involved generating approximately 180 million lines of mathematical proof code. Anthropic’s Frontier Red Team collaborated with Theorem to assign this task to Claude. The model drew on test programs from 123 libraries within Google’s OSS-Fuzz security project. This work represents a shift from traditional testing, which identifies bugs through execution, to formal verification, which proves the absence of errors through mathematical logic.

The scale of the task contrasts sharply with previous industry standards. Verifying a single small operating system kernel, known as seL4, previously required 20 person-years of specialist labor. That high cost confined formal verification to critical systems like Airbus avionics. By automating the proof generation, Claude completed the work for 123 libraries in a matter of weeks. A separate program called Rocq automatically checks the validity of these proofs, reducing the need for human reviewers in the verification loop.

The process uncovered flaws that standard testing likely missed. Claude identified hundreds of bugs, many of which appeared only under narrow conditions such as unusually large input files or specific access sequences. In one instance, the team found as many security bugs in a widely used PDF reader during one week as are typically reported for that software in an entire year. The model also detected a logical error that had remained unnoticed for more than a decade.

Anthropic emphasizes that the proofs apply only under defined constraints. The verification holds for specific processor architectures, single-threaded programs, and inputs smaller than 64 MB. These conditions do not always match real-world deployments. Additionally, the cost of running these proofs remains high, with individual tasks consuming hundreds of millions of tokens. Anthropic has not published total project costs, though data from Opus 5.5 shows improved efficiency, solving half the proofs with 40 percent fewer tokens than its predecessor.

The collaboration positions AI verification as a practical tool for software engineering. Anthropic argues that if AI can verify its own code, developers can trust AI-generated software more and spend less time on manual review. Theorem’s roadmap targets critical infrastructure components including glibc, OpenSSL, and cURL. The project shifts the focus from whether AI can write code to whether it can prove that code is safe.