LLMs are making formal software verification feasible at a much larger scale — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Claude, working with Theorem on a project the team describes as the first large-scale use of AI for production-software verification at the assembly level, produced proofs for 123 open-source libraries.
- OSS-Fuzz supplied the corpus: in a few weeks, Claude patched discovered bugs and proved 239 entry points across 123 libraries could not crash on inputs below 64 MB.
- RISC-V proofs totaled about 180 million lines of Rocq; a typical proof ran roughly 600,000 lines for binaries with a median length of 18,000 instructions.
- Claude also uncovered bugs that had eluded detection for years despite trillions of test runs.
- RISC-V results are conditional: the binaries must be single-threaded, static and fixed-memory programs running under modeled Linux conditions, so the libraries are not certified bug-free.
- Theorem is offering the pipeline for other open-source projects and publishing a Claude verification skill, while a subset of proofs and the RISC-V and OS semantics are being released for reuse and independent checking.
- Rocq checks that the formal specifications are satisfied; developers instead need to trust the RISC-V and Linux models, theorem statements and proof checker rather than manually reviewing every proof line.
Why it matters: For maintainers of the 123 libraries, Claude turns a historically years-long specialist task into a workflow that proved 239 entry points in a few weeks. The gain is scoped, reusable security evidence—not a bug-free verdict—because each demonstrated guarantee depends on fixed RISC-V/Linux assumptions and trusted formal models.
Ask SkimNews
