Claude AI Generates Formal Proofs for Software Safety
The research effort provides mathematical verification for 123 open-source libraries to prevent software crashes.
Updated on Oct. 24, 2026 in Artificial Intelligence
Live Poll
Do you trust that using AI to verify software code will make digital infrastructure more secure?
Anthropic and Theorem recently used the Claude AI model to generate formal proofs verifying no-crash behavior in 123 open-source software libraries. The research-stage project produced 180 million lines of Rocq proof code across 239 specific software entry points.
Why it matters
This application of LLMs to formal verification demonstrates an attempt to automate the complex, manual process of proving software reliability. By utilizing AI to write code proofs, the researchers aim to identify memory-safety vulnerabilities in critical open-source infrastructure.
The researchers generated 180 million lines of Rocq code, averaging 600,000 lines per proof for binaries with a median size of 18,000 RISC-V instructions. These proofs verify operation within a 64 megabyte input limit on a fixed memory budget.
The players
Anthropic
An AI safety and research company that develops the Claude series of large language models.
Theorem
A research entity focused on applying formal methods and automated reasoning to software security.
The details
The project uses Claude to write formal proofs in Rocq, a proof assistant software that converts mathematical specifications into verifiable code. By establishing strict environmental constraints—modeling programs as single-threaded static binaries on RISC-V hardware—the team reduced the complexity required for the model to successfully reason about memory safety. These proofs confirm that software will not crash when processing inputs under the specified size limits.
Timeline
The project reached completion in October 2026 after several weeks of work.
The Tech Race
While services like OSS-Fuzz have historically relied on automated fuzzing to detect existing vulnerabilities, this approach uses formal verification to guarantee software correctness. It marks a significant shift from reactive testing to proactive, AI-generated mathematical proof.
These proofs are currently in the research stage and do not yet affect standard software update cycles or end-user applications. Future implementation of these verification methods could lead to more stable open-source libraries that form the foundation of global computing infrastructure.
The takeaway
The project demonstrates that AI can handle large-scale formal verification, a traditionally arduous task for human developers. Industry observers should watch for the next phase of this research, specifically whether the team can successfully scale proof generation to multi-threaded programs or larger binary inputs.
Further reading
Explore deeper developments in Artificial Intelligence to track how models are automating complex software engineering tasks.
Source note: This article includes information reported by TokenPost.
Live Poll
Do you trust that using AI to verify software code will make digital infrastructure more secure?
