AI model automates safety checks for 123 open-source libraries

Ensuring that software functions without errors has long been a costly and labor-intensive process, typically reserved for high-stakes industries like aviation. On October 23, 2026, Anthropic announced a significant shift in this practice by using its Claude model to automatically generate mathematical proofs that verify the correctness of existing code.

The initiative, conducted in partnership with Theorem, focused on 123 open-source libraries sourced from Google’s OSS-Fuzz security project. Rather than writing new software, the Claude model translated crash tests into mathematical problems and analyzed compiled machine code. In a matter of weeks, the system produced approximately 180 million lines of proof code. According to Anthropic, an independent verification tool named Rocq automatically checks these proofs, a process that requires almost no human review.

This approach addresses a fundamental challenge in software development: the high cost of formal verification. Traditionally, proving that code is error-free under all conditions was prohibitively expensive. For instance, a team of specialists required 20 person-years to formally verify the seL4 operating system kernel. By automating this work, Anthropic aims to make formal verification as routine and accessible as standard testing.

A major benefit of the project was the discovery of numerous defects. While generating proofs, Claude identified hundreds of bugs that conventional testing had failed to catch. Many of these issues only surfaced under very specific conditions, such as unusually large input files or particular sequences of data access. In one notable instance involving a widely used PDF reader, the team found as many security bugs in a single week as are typically reported over an entire year. In another case, the model detected a logical error that had remained undetected for more than a decade.

However, the researchers note that the term "formally verified" does not equate to perfect error-free operation. The mathematical proofs hold only under strict conditions, including a specific processor architecture, single-threaded programs, and inputs smaller than 64 megabytes. Since these libraries often run on various hardware configurations, the verification remains conditional rather than absolute.

The computational cost of this automation is significant. Individual proofs can consume hundreds of millions of tokens, though Anthropic did not disclose the total project costs. Efficiency is improving, however; a smaller test run demonstrated that the newer Opus 5.5 model solves half of the proofs using roughly 40 percent fewer tokens than the previous Opus 5 version. This increased efficiency suggests that the cost of verifying AI-generated or existing code may decrease over time.

The project marks a pivotal moment for software automation. If AI can not only write code but also prove its safety within defined parameters, developers may place greater trust in automated systems. This could reduce the time spent on manual code reviews and lower overall development costs. The success of this initial effort suggests that the question for the industry is shifting from whether AI can program to whether it can guarantee the safety of that code.

Looking ahead, Anthropic expects formal verification with AI to become as practical as standard testing. Theorem’s roadmap already includes plans to extend this verification process to critical infrastructure components such as glibc, OpenSSL, and cURL. This expansion could have far-reaching implications for the reliability of the software that powers digital infrastructure.