Formal verification is a way to prove that software behaves as specified, instead of testing a few cases. On 22 October 2026, Anthropic's Frontier Red Team, working with Theorem, published research saying Claude can now do this at a much larger scale.
What Anthropic reports
Anthropic says Claude wrote formal proofs for real open-source software. The proofs cover 239 entry points across 123 OSS-Fuzz libraries. The stated property is that the code does not crash on any input under 64 MB. Anthropic says the work took a few weeks.
Claude turned existing fuzz tests into specifications. It then proved the compiled code met them. Along the way, Anthropic says, hundreds of bugs surfaced. Claude patched them and proved the fixed versions.
The authors also name the costs. Proofs can take hundreds of millions of tokens, which they call a key limitation.
The caveats matter most
Anthropic is clear that "formally verified" does not mean "bug-free." A specification can differ from real machine behavior. The proofs assume a single-threaded static binary on RISC-V under Linux. Many real deployments run on x86 or ARM with different libraries.
There is also a risk of reward hacking. Claude could pass a proof by rewriting code so it never crashes. Anthropic says it penalized patches that change behavior and audited the patches.
What it means for people who check data
The useful idea is the pairing of a claim with a checker. A proof checker accepts or rejects each step, so a reader does not have to trust the model. The same habit applies to data verification. Ask what a claim assumes, and what independent check backs it up.
The limits apply here too. A proof only covers the property that was specified. A clean result on the wrong question still misleads. The trust index and the how it works page show how we weigh evidence.
These are Anthropic's own findings. Anthropic has released a subset of the proofs and the semantics it used.