AI-generated proofs cover explicit properties, not entire libraries
Anthropic and Theorem report proving security properties for selected fuzz-test entry points across 123 open-source libraries. They use static RISC-V binaries, modeled single-threaded Linux execution, inputs below 64 MB and fixed memory assumptions. An external checker validates the proofs, and a public subset is available. The result covers those properties and assumptions rather than establishing that entire libraries are bug-free.
Anthropic and Theorem · Oct 22, 2026
Why it matters
Engineers evaluating AI-assisted verification can inspect the exact theorem, binary and machine assumptions behind a security claim.
More context
The public proof repository currently covers 61 entry points in 26 libraries and uses Rocq. Its checker needs Docker, Python dependencies and substantial memory. The collaborators say producing proofs can consume hundreds of millions of tokens; begin by inspecting one published statement and its limits.
Keep in mind
The machine model is incomplete, entry points do not cover whole libraries, and sanitizer checks miss some errors. The 123-library result is collaborator-reported; the public repository contains a smaller subset, and proofs were not rerun here.
Original read