Claude formal verification: 123 libraries proved crash-free

"Formally verified" used to mean a research team, a year, and one small program. Anthropic just made it a batch job. In AI formal verification research published October 22, its Frontier Red Team and the startup Theorem had Claude write machine-checked proofs that 123 open-source libraries cannot crash on their fuzz-test inputs. Along the way, Claude found and patched hundreds of bugs. The headline is the scale. The lesson for you is in the fine print.

What actually happened

Per Anthropic's write-up, the team took libraries from Google's OSS-Fuzz program and turned each fuzz harness into a spec: this test always exits normally, with no memory-safety errors or failed assertions, for any input under 64 MB. Claude then wrote proofs in the Rocq proof checker against the compiled RISC-V binary, not the source code.

The numbers:

• 239 entry points across 123 libraries proved crash-free under the stated conditions.
• About 180 million lines of Rocq in total. A typical proof runs around 600,000 lines.
• One proof can take hundreds of millions of tokens. Anthropic gives no dollar figure.
• On one widely used PDF reader, the team found as many security bugs in a week as usually get reported in a year.

When Claude hit a bug, it patched the code, regenerated the spec, and proved the fixed version. The team penalized patches that "fixed" a crash by gutting functionality, and audited them. Proofs and the RISC-V/Linux semantics are on GitHub. Proofs that could expose unpatched bugs are held back.

Why AI formal verification matters for your business

Anthropic says it plainly: "formally verified" does not mean "bug-free." These proofs assume a single-threaded static binary, RISC-V, musl, and Linux. Your server almost certainly runs x86 or ARM with dynamic glibc. A caller that misuses the library can still corrupt memory. The proof is exactly as strong as the spec, and no stronger.

That is the part you can use today. Within a year, vendors will put "AI-verified" on their sales pages. When one does, ask three questions: verified against what spec, on what platform, and which code paths were left out. If they can't answer, the badge is marketing.

The other signal is cost direction. Finding bugs used to be the expensive step in security. Here, a model found them as a side effect of proving the rest of the code clean. For the open-source libraries in your stack, that means more CVEs, faster. Your patch cadence has to keep up.

Sources: Anthropic, Theorem Labs on GitHub.