Anthropic says Claude proved 239 library entry points. The proof is a narrow binary, not all of software.
Anthropic says Claude proved 239 entry points in 123 libraries. The proof assumes one single-threaded RISC-V Linux binary and inputs under 64 MB.
The short version
Anthropic says Claude produced Rocq proofs, which the Rocq kernel checked, for 239 fuzz-test entry points across 123 libraries. The claim is that those programs exit without a memory-safety error or a failed assertion for every input under 64 MB, if they run as a single-threaded static binary with a fixed memory budget on RISC-V under Linux, and if the trusted semantics are right. That is not a proof that the libraries are safe on the computers people actually run.
The sentence is the assumptions
On October 22, 2026 Anthropic's Frontier Red Team posted that, with Theorem, Claude had produced formal proofs for real libraries. The concrete claim is 239 entry points across 123 open-source libraries from the fuzzing corpus. For each, the proof says the program exits normally, without a memory-safety error or a failed assertion, for every input smaller than 64 megabytes and every modeled Linux response and memory layout, if the program is a single-threaded static binary with a fixed memory budget on a RISC-V CPU running Linux, and if the semantics the proof trusts are correct. The Decoder's October 23 headline says Claude proves code safe. The headline is broader than the paragraph. This page files the paragraph.memory safety, as this paper uses the phrase.
What is being proved
Anthropic says the library and an existing fuzz test are compiled to a RISC-V executable. That executable, Theorem's model of the processor and of Linux, and the desired no-crash property become a specification. Claude writes a proof. The Rocq kernel checks it. The green region in their diagram is what is proved, conditional on the trusted parts. The proofs sit on the compiled instructions, not on a comment in the C source. Anthropic says a typical proof is about 600,000 lines for a binary whose median length is about 18,000 instructions, and that the accepted corpus is about 180 million lines. An appendix says Claude wrote on the order of a billion lines in the attempt, of which about 180 million passed the checker. Attempted lines are not accepted lines.
A public subset is smaller
NxCode says the public snapshot it looked at lists 60 entry points in 25 libraries. It treats that as a published subset, so the two counts are different artifacts. A reader who clones a repository and counts files will not reproduce 239 unless the repository is the full set Anthropic describes. This page will not scold either number. It will say which document each number comes from. The Decoder's retelling stays with Anthropic's 123 libraries and 180 million lines, and it adds the comparison to seL4, which took specialists many person-years. That comparison is the Decoder's framing of why the attempt is news. It is not a claim that these library proofs are an operating-system kernel.a release note has to name the version.
Bugs found are not the proof
Anthropic says Claude found bugs that testing had missed, patched them, and then completed the proofs. A bug found along the way is a change to the code. The proof applies to the code after that change, under the assumptions, not to every older tarball a distribution still ships. The Decoder says conventional testing had missed the bugs and that the process needs almost no human reviewers. Almost no reviewers is the outlet's compression of Anthropic's autonomous-proof claim. The tools and the semantics, Anthropic says, were built by people at Theorem. The model did not invent RISC-V.Linux's merge window is a different artifact.
What the proof refuses to cover
Another processor. More than one thread. An input at or above 64 megabytes. A dynamic linker. A memory budget other than the one modeled. A property other than normal exit without a memory-safety fault or a failed assertion. Anthropic says expanding the semantics to more architectures is future work, not this result. A library that is safe inside those walls can still be dangerous in the build a server actually runs. Formal verification, Anthropic also says, can prove less about security than a reader hears in the phrase proved safe. This page keeps that warning next to the number 239.
The appendix's other experiment
The appendix describes a separate attempt: a Python interpreter fragment written in Rust and checked with Verus, against a formal semantics the authors tested empirically against CPython with a large set of programs. That experiment is not the 239 library proofs. A reader who merges them will think one project both compiled C libraries to RISC-V and implemented Python. They are both in the same PDF, and they are different claims. This page's row is the library proofs. The interpreter is noted so it is not silently added to the 239.what a training cutoff does not mean.
Who is in scope
Maintainers of the 123 libraries, if their shipped binaries match the proved configuration, which most will not. Researchers who want the assumptions in one place. Anyone who read a headline that said the code was proved safe and needed the boundary. People running ordinary x86 servers are outside the proved configuration even if the source tree is the same.an evaluation a second lab can run.
What stays on the record
October 22, 2026 Anthropic research note. 239 entry points, 123 libraries, about 180 million accepted lines of Rocq, typical proof about 600,000 lines, median binary about 18,000 instructions. Assumptions: single-threaded static RISC-V Linux binary, fixed memory budget, inputs under 64 MB, trusted semantics. Rocq checks the proof. Bugs found and patched before the proof, on Anthropic's account. Decoder's October 23 retelling, including its hundreds-of-bugs wording. NxCode's public subset of 60 entry points in 25 libraries. Not a general safety proof.
What would move the row
A release of the full 239, or a proof for a second architecture, would be a later document. A library maintainer who says the proved patch is the version they ship would connect the note to a tarball. Until then the record is the research note, one news retelling, and a count of a smaller public snapshot.the advisory is the story.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
Does this mean the libraries cannot crash?
Only inside the assumptions. Anthropic names a single-threaded static RISC-V binary, a fixed memory budget, Linux, and inputs under 64 megabytes. A multi-threaded build, another processor, or a larger input is outside the sentence.
Who checks the proof?
Anthropic says the Rocq kernel checks the proof against a specification built from the binary, Theorem's semantics, and a no-crash property. A kernel accepting a proof is not a human reading every line. Anthropic says the accepted proofs are about 180 million lines of Rocq.
Did Claude find bugs first?
Anthropic says that while writing the proofs Claude found bugs that had survived a large number of tests, patched them, and then proved the no-crash property. The Decoder says hundreds of bugs. This page keeps hundreds as the Decoder's word and the discovery claim as Anthropic's.
Is the public proof the same size as the claim?
NxCode says a public snapshot it inspected lists proofs for 60 entry points in 25 libraries, and that this is a subset rather than the full 239. A subset on a download page and a total in a research note are both on the record. They are not the same count.
Does the proof show the program finishes?
Anthropic's note is about exiting without memory-safety errors or failed assertions under the assumptions. A separate commentary has stressed that these proofs are not a promise the program always terminates. This page does not upgrade the note into that promise.