Anthropic Frontier Red Team: Claude, working with Theorem, formally proved no-crash properties at the RISC-V assembly level for 239 entry points across 123 OSS-Fuzz libraries, about 180M lines of Rocq

Background and pain points

AI is getting better at finding bugs and chaining them into exploits, while tests and fuzzing can only try more inputs and never prove the absence of bugs. Formal verification gives a mathematical guarantee but has been extremely expensive: the seL4 kernel took 2 person-years to build and 20 to verify, so it was used only for code like avionics. Anthropic’s Frontier Red Team set out to show that LLMs now make formal verification practical for large amounts of real production code.

How it works

The team reused developer-written fuzz entry points from Google’s OSS-Fuzz as specs: each is compiled to a RISC-V executable and combined with Theorem’s RISC-V instruction semantics and Linux process semantics to produce the theorem that no input under 64 MB causes a crash. Claude writes the proof and the Rocq kernel checks it. Because OSS-Fuzz builds with AddressSanitizer, no crash also means no memory error that ASan would catch. When a proof got stuck on a real bug, Claude patched the code, regenerated the spec and proved it again. The only trusted pieces left are the RISC-V and Linux models, the theorem statements and the Rocq checker.

Benchmarks

Anthropic’s figures: over a few weeks Claude proved 239 entry points across 123 libraries, about 180 million lines of Rocq, with a typical proof around 600,000 lines. It found hundreds of bugs, including an out-of-bounds write that needed a roughly 20 MB input, and rediscovered CVE-2026-56132 in libexpat. On 36 small tasks Opus 5.5 proved half with about 40% fewer tokens than Opus 5, and a shared workspace cut tokens another 17%. Side experiments proved the Linux kernel’s zlib matches RFC 1950 using Claude-written x86 semantics, and a Rust-based Python interpreter matched CPython on 99.95% of 1.7 million random programs.

Getting started

Know the limits: proofs cover single-threaded static RISC-V binaries with musl, while most deployments are x86 or ARM with glibc; ASan misses some memory errors, and one proof can take hundreds of millions of tokens. To try it, re-check the published proofs in theorem-labs/oss-verify-proofs on GitHub, use the Claude skill Theorem published to set up your own project, or ask Theorem to run the same pipeline for an open-source project.