Claude Automates Software Safety Checks at Scale

180,000,000. That is the number of lines of mathematical proof code generated by Anthropic’s Claude over just a few weeks. The massive output was not new software, but a rigorous mathematical validation of existing code. This work represents a significant shift in how the industry approaches software reliability.

The effort was led by Anthropic’s Frontier Red Team in partnership with Theorem. Instead of writing new applications, Claude took on the task of formally verifying existing software. The model drew on test programs from 123 open-source libraries within Google’s OSS-Fuzz security project. The goal was to mathematically prove that this code behaves correctly under defined conditions.

Formal verification has historically been prohibitively expensive and slow. For context, a team of specialists required 20 person-years to verify the small operating system kernel seL4. Because of this cost, the method was typically reserved for critical systems like Airbus avionics. By using Claude, Anthropic compressed this timeline from years to weeks, demonstrating that AI can handle the heavy lifting of complex proof generation.

The process was described as fully autonomous, though it relied on human-built tools. Claude translated crash tests into mathematical problems and analyzed compiled machine code. It then produced extensive proofs that were checked by an independent verification program called Rocq. According to Anthropic, this workflow required almost no human reviewers, marking a major step toward automated software assurance.

The most immediate benefit of this process is the discovery of hidden defects. While generating proofs, Claude identified and fixed hundreds of bugs. Many of these errors would have been missed by conventional testing because they only appear under specific conditions, such as unusually large input files or specific sequences of memory access. The AI was able to pinpoint these edge cases that traditional methods often overlook.

The impact was stark in specific cases. When verifying a widely used PDF reader, the team found as many security bugs in one week as are typically reported for that software in an entire year. In another instance, Claude caught a logical error that had remained undetected for over a decade. Even when a complete proof failed, the attempt still surfaced real, actionable bugs, proving the value of the exercise.

However, there are strict limitations to what this achievement means. Anthropic notes that formal verification does not equate to being error-free in every possible scenario. The proofs only hold under specific constraints, including a particular processor architecture, single-threaded programs, and inputs under 64 MB. In real-world practice, these libraries often run on diverse hardware and configurations, meaning the verification is context-dependent.

The computational cost remains a factor. Individual proofs can consume hundreds of millions of tokens, and Anthropic has not disclosed the total project costs. Nevertheless, the efficiency is improving. A smaller test run showed that proofs are reusable, and newer models are more efficient. Opus 5.5, for example, solves half the proofs using about 40 percent fewer tokens than its predecessor, Opus 5. This trend suggests that the cost barrier may continue to lower over time.

This project shifts the conversation in software development. If AI can write code and guarantee its correctness under defined conditions, the focus moves from basic programming capability to safety assurance. Anthropic expects this type of formal verification to become as routine as standard testing. Partner Theorem already has a roadmap targeting critical infrastructure components like glibc, OpenSSL, and cURL, signaling a broader push toward automated safety in the tech stack.