Anthropic partners with Theorem to use Claude for formal assembly‑level security proofs in over 100 open‑source libraries
anthropic claude open-source
| Source: Techmeme | Original article
Anthropic has teamed with Theorem to employ its Claude AI for formal, assembly‑level security proofs across more than 100 open‑source libraries.
Anthropic announced a partnership with verification specialist Theorem to put its Claude model to work on formal security proofs for open‑source software. In research released on 22 October 2026, the two teams used Claude to translate fuzz‑test cases from Google’s OSS‑Fuzz suite into formal specifications and then generate machine‑checked proofs that the compiled RISC‑V binaries cannot crash on any input under 64 MB. The effort covered 239 entry points across 123 libraries – more than a hundred real‑world projects – and produced roughly 180 million lines of proof code checked by the Rocq kernel.
The achievement matters because it demonstrates that large language models can move beyond code generation to rigorous verification at the assembly level, a domain traditionally reserved for hand‑crafted proofs. Claude not only produced the proofs but also uncovered hundreds of bugs that had survived extensive conventional testing, highlighting a new avenue for strengthening the software supply chain. By automating the creation of formal guarantees, the approach could reduce reliance on costly manual audits and make high‑assurance security more accessible to the open‑source ecosystem.
Looking ahead, the community will be watching whether the workflow can be scaled to larger codebases and other architectures, and how quickly developers will integrate such LLM‑driven verification into continuous‑integration pipelines. Anthropic’s Frontier Red Team plans to publish further results, and industry observers expect additional collaborations that could extend formal guarantees to critical infrastructure and commercial products. The partnership signals a shift toward AI‑augmented formal methods, a trend that could reshape how software reliability is assured in the coming years.
Sources
Back to AIPULSEN