Algorithmic Assurance: AI Generates Mathematical Proofs for Legacy Code

On October 23, 2026, Anthropic disclosed that its Claude model has successfully generated millions of lines of mathematical proof code to verify the correctness of existing software libraries. This development marks a significant shift in software engineering, moving beyond the generation of new code to the automated certification of legacy systems. The initiative, conducted in partnership with Theorem, suggests that artificial intelligence can now handle the rigorous, yet historically expensive, task of formal verification.

The central claim of the project is that AI can replace human specialists in the labor-intensive process of proving that software runs error-free. By leveraging Claude, the team produced approximately 180 million lines of proof code in a matter of weeks. This work covered 123 open-source libraries sourced from Google’s OSS-Fuzz security project. The model did not write new applications but instead analyzed compiled machine code, translated crash tests into mathematical problems, and produced extensive proofs to validate the existing logic.

The implications for software automation are substantial, as formal verification has traditionally been reserved for high-stakes environments like avionics due to its high cost. Previously, verifying a small operating system kernel like seL4 required twenty person-years of specialist labor. By automating this process, Anthropic aims to make such rigorous checks as routine as standard testing. The researchers describe the process as "fully autonomous," meaning that once the formal foundations were established by humans at Theorem, the model handled the proof generation with minimal human intervention.

During the verification process, the model identified hundreds of bugs that conventional testing methods had overlooked. These errors often remained hidden because they only manifested under specific conditions, such as unusually large input files or particular sequences of data access. In one notable instance, the team found as many security flaws in a single week within a widely used PDF reader as are typically reported in an entire year. In another case, a logical error that had persisted for over a decade was detected. Even when a complete proof could not be finalized, the attempt successfully surfaced these latent defects.

However, the project comes with significant caveats regarding the scope of the verification. Anthropic clarifies that "formally verified" does not equate to "error-free" in all contexts. The generated proofs are valid only under strict conditions, including a specific processor architecture, single-threaded execution, and input files smaller than 64 MB. Since these libraries often run on diverse hardware and in multi-threaded environments, the proofs provide a high level of confidence but not universal guarantees for every possible deployment scenario.

The computational cost of this automated verification remains a factor to consider. Individual proofs can require hundreds of millions of tokens, and while Anthropic has not disclosed the total project costs, the efficiency of the process is improving. A smaller test run indicated that proofs are reusable and that newer model versions are more efficient. Specifically, Opus 5.5 solved half of the proofs using approximately 40 percent fewer tokens than the previous Opus 5 version, suggesting a trajectory toward greater cost-effectiveness.

Independent verification of the generated proofs is handled by a program called Rocq, which automatically checks the mathematical consistency of the results. This layer of validation is critical to ensuring that the AI's output is not just plausible, but rigorously correct. The integration of such tools into the development pipeline could reduce the time developers spend on manual code review, thereby lowering costs across the entire software lifecycle.

Looking ahead, Theorem’s roadmap targets critical infrastructure libraries such as glibc, OpenSSL, and cURL. Anthropic expects that formal verification using AI will soon become as practical as standard testing procedures. If successful, this approach could transform the industry by allowing developers to trust AI-generated software more deeply, shifting the primary question from whether AI can code to whether it can prove that its code is safe under defined parameters.