The proofs were written at the assembly-code level for RISC-V processors, the compiled instructions a chip actually runs.
Anthropic says Claude showed that 239 entry points across the libraries will never crash on any input under 64 MB, and patched bugs it found along the way.
Anthropic describes this as, to its knowledge, the first large-scale effort to use AI to formally verify production software at that level.
Formal verification, which mathematically proves that code follows defined rules, has historically been practical only for mission-critical systems because of the effort involved.
Anthropic cites the seL4 kernel, which took two person-years to build and 20 to verify.
Here, Claude turned existing fuzz tests, which bombard code with generated inputs, into specifications, so a completed proof means fuzzing that entry point can find no more crashes.
The company says the process uncovered hundreds of bugs, including ones fuzzers would be unlikely to catch, such as a memory write error triggered only by roughly 20 MB inputs with a rare layout and a rediscovery of CVE-2026-56132 in libexpat.
Anthropic stresses that "formally verified" does not mean bug-free. The proofs assume a single-threaded static binary with a fixed memory budget on a RISC-V CPU running Linux, while most users run these libraries on x86 or ARM. They also cover only the tested entry points, not whole libraries. Remaining risk sits in the trusted components: the RISC-V and Linux models, the theorem statements, and the Rocq proof checker. The proofs total about 180 million lines of Rocq, and producing them can take hundreds of millions of tokens, though Anthropic says newer models need fewer and that letting agents reuse each other's proofs cut token use by 17% in one test. It is publishing a subset of the proofs and the trusted components, and Theorem is offering to run the pipeline for other open-source projects and has released a Claude skill for developers.