Anthropic says Claude proved properties in 123 open-source libraries
Original title: LLMs are making formal software verification feasible at a much larger scale
Proofs cover single-threaded RISC-V static binaries and inputs under 64 MB; a subset and a Claude skill are being published.
Anthropic, with collaborator Theorem, says Claude produced formal proofs of security properties at RISC-V assembly level for 123 open-source libraries. The claimed guarantee is normal exit without memory-safety errors or failed assertions for every input under 64 MB and every modeled Linux response, under stated assumptions. Anthropic says Claude also found long-standing bugs while writing proofs. A subset of proofs, RISC-V and OS semantics, and a Claude skill are being published; Theorem offers the same pipeline to other projects.
Unconfirmed: Bug findings and the scale of the proofs are Anthropic's claims; independent verification is not reported here.
- Assumptions include a single-threaded static binary and a fixed memory budget.
- Anthropic calls this the first large-scale AI verification of production software at assembly level.
- Other experiments covered zlib, a Rust Python interpreter, and older bugs.
- The post says proofs were many times larger than code in past human efforts such as seL4.
From the original article
14 min read at anthropic.com
Keane Lucas, Aengus Lynch, Luna Tong, Luke Champine, Zygi Straznickas, Dino Distefano, Harshikaa Agrawal Evyatar Ben Asher, Ava Howell, Sami Jawhar, Chris Rohlf, Nickolai Zeldovich Most software is riddled with bugs, which AI can now easily find and exploit . The standard defenses against cyberattacks are no longer sufficient: we need ways to prove that code is free of bugs.
By johnbender · anthropic.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.