Anthropic's Claude Automates the Proving of Code Safety

Software engineers have long considered the mathematical proof of code correctness to be a luxury reserved for critical systems like aviation controls. This status quo is changing rapidly in the laboratory.

Anthropic has demonstrated a new capability where its Claude model automatically generates formal proofs for existing software. The initiative was developed in collaboration with a partner firm called Theorem. The system does not write new applications from scratch. Instead, it takes established code and mathematically verifies its behavior.

The scale of this effort is significant. Over a period of just a few weeks, the model produced approximately 180 million lines of proof code. This work covered 123 open-source libraries that were part of a security project at Google. The process relies on translating crash tests into mathematical problems. It then analyzes the compiled machine code to produce extensive proofs. An independent program called Rocq checks these proofs automatically. The researchers describe the workflow as fully autonomous, even though the underlying formal foundations were created by human engineers.

One of the most striking outcomes of the project is the discovery of hidden flaws. Claude identified hundreds of bugs during the verification process. Many of these defects would have likely passed conventional testing. They often only appear under very specific conditions, such as unusually large input files or specific sequences of data access. In one instance involving a widely used PDF reader, the team found as many security bugs in a single week as are typically reported for that software in an entire year. In another case, the model caught a logical error that had remained undetected for more than a decade.

Despite these successes, the technology has strict practical limits. Anthropic notes that a formal verification does not guarantee that software is entirely error-free. The proofs only hold under specific conditions. These include a specific processor architecture, single-threaded programs, and inputs smaller than 64 megabytes. In real-world scenarios, these libraries often run on different hardware and configurations. The computational cost is also high, with individual proofs requiring hundreds of millions of tokens. The company has not disclosed the total cost of the project, but smaller test runs show that proofs are reusable and that newer models are becoming more efficient. For example, the Opus 5.5 model solves half the proofs with about 40 percent fewer tokens than its predecessor.

The broader implication is a shift in how the industry views AI automation. If an AI can write code and also prove its safety under defined conditions, the focus moves from basic coding capability to verifiable correctness. This could reduce the time developers spend on manual review. Anthropic expects this method to become as routine as standard testing in the near future. Theorem has already outlined a roadmap that includes critical infrastructure libraries such as glibc, OpenSSL, and cURL. This development suggests that the next leap in software automation may come from the ability to mathematically guarantee safety.

The project highlights a potential path forward for trust in AI-generated software. By automating the expensive and complex task of formal verification, companies may be able to deploy AI-written code with greater confidence. The efficiency gains observed in recent model versions indicate that the cost of this verification will likely decrease over time. This sets the stage for a future where mathematical proof becomes a standard part of the development cycle rather than a specialized exception.