AI · Published
Anthropic published a research post on 22 October 2026 describing work done with the verification company Theorem. In the post, Claude wrote formal proofs about compiled code from 123 open-source libraries that are tested by the OSS-Fuzz fuzzing project. The Rocq proof checker verified the proofs, and they cover 239 entry points across those libraries.
The method turned each library’s existing fuzz tests into formal specifications over RISC-V compiled binaries. Claude wrote the proofs against those specifications, and Theorem’s model of RISC-V and Linux semantics was used to describe how the binaries behave. Anthropic says the proofs assume a single-threaded static binary with a fixed memory budget, running on a RISC-V processor under Linux.
The headline result is that Claude first patched the bugs it found and then proved there are no crashes on any input up to 64 MB for the 239 entry points. The post says the proof attempts discovered hundreds of bugs. Examples given include a counter underflow in a decompressor that caused an out-of-bounds write, an out-of-bounds read in an image decoder, a rediscovery of CVE-2026-56132 in libexpat, and a logic bug in which a file-restriction feature is not always enforced. The post also describes a logic bug that had been latent for more than a decade, and minor zlib corner-case bugs that the post describes as unexploitable.
Anthropic lists its own caveats. It says that formally verified does not mean bug-free, because bugs can arise where a specification or machine model differs from reality. The proofs cover musl-linked static RISC-V builds, not the x86 or ARM glibc builds that most people run. The post also says that AddressSanitizer does not catch every memory-safety bug, that not every entry point was verified, and that the proofs do not cover a caller misusing a library. Producing the proofs took hundreds of millions of tokens in some cases.
The post gives no severity ratings or CVSS scores for the bugs it found, so their real-world impact cannot be judged from it. It also does not clearly say whether the OSS-Fuzz bugs were reported upstream. The authors say they are not releasing proofs that might reveal an unpatched bug. For the zlib corner cases, the post says filings to the upstream project are in progress.
Why it matters: the work shows an AI system producing machine-checked proofs about real, widely used code at scale, which is a different claim from finding bugs with a fuzzer. The post reports cost figures that point downward. It says Opus 5.5 used about 40% fewer tokens than Opus 5 on 36 small tasks, and that reusing proofs across projects cut token use by 17% in a shared-workspace test. Those numbers come from Anthropic and have not been independently checked.
The stated next steps include generating accurate x86 semantics, standardising semantics for more architectures and operating-system processes, and publishing a subset of the proofs and trusted components. Anthropic says it will offer the Theorem pipeline to other projects. Two AI news sites, AIWeekly and FourWeekMBA, also covered the work. Their coverage repeats the headline figures and was not used to check them.
Several things are unknown. The post does not say how many of the hundreds of bugs are exploitable, and it does not say how upstream maintainers responded. It is also unclear how well the results will carry over to x86 and ARM builds, which the post names as future work. Readers should treat this as a company research result, with its method and scope set out by the company, until outside researchers have examined the proofs.
Sources
This article was researched and written by an AI system from the sources listed above, under SavvyToolz's sourcing and freshness rules. Read the editorial policy.