By Hamza Faisal

Anthropic says Claude can now do something that once took formal-verification experts years of painstaking work: write mathematical proofs that real production software cannot crash. In research published October 22 on its Frontier Red Team blog, the company reports that Claude — working with the formal-verification startup Theorem — produced machine-checked proofs covering 239 entry points across 123 open-source libraries, finding and patching hundreds of bugs along the way.

From fuzz tests to mathematical proof

The project started from Google's OSS-Fuzz corpus, the continuous fuzzing programme that has been hammering open-source C and C++ code since 2016. Fuzzing works by throwing semi-random inputs at an entry point until something crashes — but it can never try every input. Claude's job was to turn each existing fuzz test into a theorem: for every input under 64 MB, this entry point exits normally, with no memory-safety error and no failed assertion.

Claude then proved each theorem against the compiled RISC-V machine code, using Theorem's formal models of the RISC-V chip design and of a Linux process. The Rocq proof checker verifies each proof mechanically — a file a checker either accepts or rejects, rather than an argument a reader has to take on faith. Anthropic says virtually none of the proof needs human review: you only have to trust the RISC-V model, the Linux process model, the theorem statements and the checker itself.

A billion lines of proof — and bugs nobody knew about

The finished proofs run to roughly 180 million lines of Rocq, with a typical proof around 600,000 lines long covering a binary with a median length of 18,000 RISC-V instructions — all produced in a few weeks. Along the way, Claude found bugs that fuzzing had missed despite trillions of runs: a rediscovery of CVE-2026-56132 in libexpat, a decompressor flaw that fires only on a rare 20 MB input, and, in one PDF reader, as many security bugs in a week as are typically reported in a year.

Each time Claude found a flaw, it patched the code and proved the patched version crash-free. A subset of the proofs, along with the RISC-V and operating-system semantics underneath them, is being published on GitHub so the checking doesn't depend on trusting the lab — though Anthropic is withholding proofs that could reveal unpatched bugs.

Why it matters now — and where the limits are

The result lands in a week of AI-found bugs: an OpenAI model helped surface the critical PHP2Shell flaws, and Anthropic said Claude stripped unneeded code out of OpenSSH earlier in the week. Anthropic opens its post with the stakes stated plainly: "Most software is riddled with bugs, which AI can now easily find and exploit," it writes — "we need ways to prove that code is free of bugs."

There are limits. "Formally verified" does not mean bug-free: the proofs cover single-threaded, statically linked RISC-V binaries against musl libc, while most people run these libraries on x86 or ARM, dynamically linked against glibc. Formal verification was historically so expensive it was impractical for most projects — the seL4 kernel took two person-years to build and twenty to verify. Anthropic's bet is that LLMs collapse that cost, predicting formal verification "will soon be as practical as fuzzing is today." Theorem is offering to run the pipeline for other open-source projects and has released a Claude skill to go with it.

By Hamza Faisal