AI Proves Software Safety Automatically

The process of verifying that software operates without error under all possible conditions has historically been an expensive and labor-intensive endeavor. A team of specialists once required twenty person-years to formally verify the seL4 operating system kernel, a task that was previously reserved for critical systems such as avionics in aircraft.

On October 23, 2026, Anthropic announced a significant advancement in this field through a partnership with Theorem. The companies utilized the Claude model to mathematically verify existing software code. The system generated approximately 180 million lines of proof code in just a few weeks. This effort covered 123 open-source libraries from Google's OSS-Fuzz security project.

The approach differs from traditional AI coding tasks where the model creates new code from scratch. Instead, Claude translated crash tests into mathematical problems and analyzed the compiled machine code. It then produced extensive proofs that were automatically checked by an independent verification program called Rocq. Anthropic describes the process as fully autonomous, noting that it requires almost no human reviewers. The underlying formal foundations and tools were developed by human experts at Theorem.

During the verification process, the model identified hundreds of bugs that conventional testing methods had missed. Many of these errors only surfaced under specific conditions, such as unusually large input files or particular sequences of memory access. The researchers noted that these issues would likely have slipped past standard testing protocols.

The efficacy of the system was highlighted during the verification of a widely used PDF reader. The team discovered as many security bugs in one week as are typically reported for that software in an entire year. In another instance, Claude detected a logical error that had remained undetected for more than a decade. Anthropic states that even when a complete proof fails, the attempt often reveals real bugs in the software.

The company acknowledges that formal verification does not guarantee that software is entirely error-free. The proofs hold only under strict conditions, including a specific processor architecture, single-threaded programs, and input sizes under 64 MB. In practical applications, these libraries often run on diverse hardware and configurations, which presents a limitation to the current method.

The computational cost of generating these proofs is substantial. Individual proofs can require hundreds of millions of tokens, though Anthropic has not disclosed the total cost of the project. However, the process offers potential savings by allowing developers to trust AI-generated software more confidently. This reduces the time spent on manual review. A smaller test run indicated that proofs are reusable, and newer models like Opus 5.5 are more efficient, solving half the proofs with about 40 percent fewer tokens than the previous version.

This project strengthens the argument for broader software automation. If AI can write code and guarantee its correctness under defined conditions, the focus shifts from basic programming capabilities to safety assurance. Anthropic expects formal verification with AI to become as routine as standard testing. Theorem's roadmap already targets critical infrastructure components such as glibc, OpenSSL, and cURL for similar verification efforts.