AI Proves 180 Million Code Lines

Anthropic has demonstrated that its Claude model can mathematically verify the correctness of existing software code, a capability that traditionally required significant human expertise. The company positions this development as a critical step toward increasing trust in AI-generated software. While the technology offers a new path for assurance, Anthropic emphasizes that the proofs operate within strict technical boundaries rather than guaranteeing universal error-free performance.

The efficiency gains are substantial when measured against historical baselines. Verifying seL4, a small operating system kernel, previously demanded approximately 20 person-years of specialized labor. In contrast, Claude generated around 180 million lines of mathematical proof code in a matter of weeks. This output was derived from test programs associated with 123 open-source libraries in Google's OSS-Fuzz security project. The scale of this automated process contrasts sharply with the limited adoption of formal verification in sectors like aviation, where costs previously restricted the practice to extreme cases.

The initiative was executed by Anthropic's Frontier Red Team in collaboration with Theorem. The model did not write new software. Instead, it transformed crash tests into mathematical problems and analyzed compiled machine code to construct extensive proofs. A separate verification tool named Rocq automatically checks these proofs. The researchers describe the generation process as fully autonomous, although they note that the underlying formal foundations and tools were developed by human experts at Theorem.

The verification process yielded immediate practical benefits by uncovering hidden defects. Claude identified hundreds of bugs during the work and was authorized to fix them. Many of these issues likely would have escaped conventional testing because they manifest 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 in a single week as are typically reported in a full year. In another case, the model detected a logical error that had remained unnoticed for over a decade.

Financial considerations remain a significant factor for broader adoption. Individual proofs consume hundreds of millions of tokens, and Anthropic has not disclosed the total project costs. However, the company argues that verifying AI code can reduce overall development expenses by minimizing the time developers spend on manual review. A smaller test run suggested that proofs can be reused, and efficiency is increasing. Specifically, Opus 5.5 solves half the proofs while using approximately 40 percent fewer tokens than its predecessor.

The company maintains that formal verification does not equate to being error-free. The proofs are valid only under defined conditions, including a particular processor architecture, single-threaded programs, and inputs smaller than 64 MB. Real-world deployments often run on different hardware configurations, which limits the scope of these guarantees. Despite these constraints, Anthropic expects AI-driven verification to become as practical as standard testing.

The project shifts the focus from whether AI can program to whether it can prove its code is safe. Theorem's roadmap already targets critical infrastructure components such as glibc, OpenSSL, and cURL. This strategic move aims to establish a reliable foundation for future software development pipelines.