Anthropic Says Claude Formally Verified More Than 100 Open-Source Software Libraries

Picture the kind of painstaking work that normally takes a team of specialist mathematicians months to complete. Anthropic says its AI model Claude has just done a version of that — at scale — for more than a hundred pieces of software that millions of people and organisations use every day.

Anthropic’s Frontier Red Team, working alongside formal-methods company Theorem, used Claude to produce formal mathematical proofs for security properties across 123 open-source libraries drawn from Google’s OSS-Fuzz corpus. The project covered 239 fuzzing entry points and, according to Anthropic, generated around 180 million lines of Rocq proof code in the process.

That is not a typo. One hundred and eighty million lines.

What Formal Verification Actually Means

Most software security testing works by throwing things at a programme and seeing what breaks. Fuzz testing — the approach used by OSS-Fuzz — does exactly that, running software with enormous numbers of inputs to hunt for crashes or unexpected behaviour. It’s useful, but it can’t test every possible input. There will always be gaps.

Formal verification takes a different approach. Instead of testing many inputs, it uses mathematical proofs to establish that a defined property holds across all inputs covered by the formal model. Think of it as the difference between checking whether a bridge holds up under a thousand different lorries, and mathematically proving the conditions under which it will or won’t collapse.

The catch has always been cost. Formal verification is expensive, specialist work. Applying it to large, real-world codebases has historically been impractical for most organisations. That’s the barrier Anthropic says it’s working to lower.

What Claude and Theorem Actually Did

The proofs weren’t applied to tidy, readable source code. They were applied to compiled RISC-V assembly — the actual machine-level instructions a processor executes. The median binary contained around 18,000 RISC-V instructions, and a typical proof ran to about 600,000 lines of Rocq code. The proofs weren’t just taken on Claude’s word. They were checked using the Rocq proof assistant, a machine-checked validation tool. That distinction matters: Claude generated the proofs, but Rocq independently verified them, providing a layer of assurance that goes beyond trusting the AI’s output.

The team proved that all 239 entry points would not crash on any input up to 64 MB in size, within the conditions of the formal model. And along the way, Claude helped identify bugs that had survived extensive prior testing without being caught.

A First — With Caveats

Anthropic describes this as, to its knowledge, the first large-scale use of AI to apply formal verification to production software at assembly-code level. That claim hasn’t been independently verified, and it’s worth treating it as Anthropic’s own characterisation rather than an established consensus.

The security guarantees themselves also come with limits. Formal proofs apply only to the specific properties, assumptions, and execution environment that the model describes. They don’t mean the libraries are free from every possible vulnerability. A proof that software won’t crash on inputs up to 64 MB says nothing about, say, a logic flaw that only appears under a different set of conditions. Conventional security testing, vulnerability management, and human review still have a role to play.

On top of that, the findings haven’t been through independent peer review, based on available sources. Anthropic published the work through its Frontier Red Team in October 2026.

So this is a promising set of results from a well-resourced team, not yet a settled chapter in computer science.

Why Open-Source Security Matters

The libraries involved aren’t obscure. OSS-Fuzz is a Google-backed continuous fuzzing programme that covers widely used open-source software — the kind of code that sits inside operating systems, web servers, developer tools, and countless applications. When a vulnerability turns up in one of these libraries, the ripple effects can be wide.

At the same time, the argument Anthropic is making is that AI can reduce the expert time needed to apply formal verification, making mathematically checked guarantees feasible for far larger software projects than was previously practical. If that holds up under scrutiny, it could change how the industry thinks about security assurance for the open-source code that underpins so much of the modern internet.

But reproducibility, maintenance burden — keeping proofs valid as code evolves — and the availability of skilled reviewers to oversee the process all remain live questions. Scaling this beyond a research project into routine practice is a different challenge from demonstrating it’s possible in the first place.

What This Means for Kent Residents

There’s no direct Kent-specific angle here — Anthropic hasn’t announced any local deployment, and no Kent organisation is known to have been involved in the project. But the broader picture is relevant to anyone who relies on public digital services. Organisations like Kent County Council, Medway Council, and NHS Kent and Medway ICB all depend on open-source software to run services that residents use daily. If AI-assisted formal verification matures into a reliable, affordable tool, it could eventually offer stronger security assurances for the software those organisations run — though that outcome depends on the techniques proving reproducible and cost-effective well beyond this initial research phase.