Claude did not write new code. It translated crash tests into mathematical problems, analysed the compiled machine code and produced proofs that an independent program called Rocq automatically checks. Anthropic says the process needs almost no human reviewers, though the researchers describe it as "fully autonomous" with the caveat that the tools and formal foundations were built by humans at Theorem.

The proof work surfaced real bugs. While verifying a widely used PDF reader, the team reportedly found as many security bugs in one week as are typically reported for that software in an entire year, and in another case Claude caught a logical error that had gone undetected for over a decade. Many of the bugs would evade conventional testing because they only appear under specific conditions, such as unusually large input files or a particular sequence of accesses.

The guarantees are narrower than the word "verified" suggests. Anthropic acknowledges the proofs hold only under strict conditions: a specific processor architecture, single-threaded programs and inputs under 64 MB, while real deployments use different hardware and configurations. Compute costs are also steep, with individual proofs consuming hundreds of millions of tokens, and Anthropic does not report the total project cost. A smaller test run suggests proofs are reusable, and newer models are more efficient, with Opus 5.5 solving half the proofs using about 40 percent fewer tokens than Opus 5.

Anthropic expects AI-assisted formal verification to become as practical as standard testing, and Theorem's roadmap targets critical infrastructure libraries including glibc, OpenSSL and cURL.