Anthropic's Claude Creates Proofs to Validate Software Safety
The AI company is using its model to mathematically verify code, generating millions of lines of proof code to identify hidden bugs and ensure safety.
By Kavya Kapoor ·
Software safety has long been a difficult problem because mathematically proving code runs without errors is historically too costly for most projects. Anthropic announced on October 23, 2026, that its Claude model has taken on the task of verifying software safety by generating the necessary mathematical proofs.
The effort was carried out by Anthropic's Frontier Red Team in partnership with Theorem. In just a few weeks, the system generated approximately 180 million lines of proof code covering 123 open-source libraries. This scale of output was previously unthinkable, as verifying a single operating system kernel once required 20 person-years of specialist labor.
Claude achieved this by converting crash test scenarios into mathematical problems and analyzing compiled machine code to produce proofs. These proofs are then automatically checked by a verification program called Rocq, a process Anthropic describes as fully autonomous. The tools and formal foundations involved were, however, built by human engineers at Theorem.
While creating these proofs, the model identified hundreds of bugs that conventional testing had missed. Many of these flaws only appear under specific conditions, such as unusually large input files. During one week spent verifying a PDF reader, the system uncovered as many security bugs as are typically reported for that software in an entire year, and in another case, it found an error that had been undetected for over a decade.
Anthropic notes that while the model can generate proofs, it cannot guarantee an application is entirely error-free. The proofs only hold under strict conditions, such as using a specific processor architecture, single-threaded programs, and inputs under 64 MB. Additionally, the compute cost is steep, with individual proofs consuming up to hundreds of millions of tokens.
The project demonstrates the potential for software automation to advance significantly if AI can both write code and prove its safety. Anthropic expects formal verification with AI to become as routine as standard testing. Theorem's roadmap currently targets critical infrastructure components such as glibc, OpenSSL, and cURL.
Efficiency is improving with newer models. Anthropic states that Opus 5.5 can solve half the proofs with about 40 percent fewer tokens than Opus 5. Proofs are also reusable, which could reduce costs for future projects involving the same code.
Source: The Decoder
