Claude turns fuzz tests into no-crash proofs for 123 OSS libraries
Anthropic open-sources the proofs and a free OSS Verify program — but they cover RISC-V/musl binaries, not your x86 build. Plus an AI SDK tool-call guard.
Copy markdown
239 fuzz tests, now machine-checked no-crash proofs
Anthropic and Theorem had Claude convert 239 fuzz entry points across 123 OSS-Fuzz libraries into formal Rocq proofs that the code simply never crashes — roughly 180M lines of generated proof passed the checker. For once, “it fuzzes clean” is upgraded to “proven not to crash on any input under 64MB.”
Get your own library proved, for free
Theorem's OSS Verify takes qualifying open-source projects via a pull request (add project.yaml, Dockerfile, build.sh); bugs are disclosed privately to maintainers with no deadline. Anthropic also open-sourced a subset of the proofs, its RISC-V and OS semantics, a 36-task benchmark, and a Verus-verified Python interpreter written in Rust.
Read the fine print before you trust the checkmark
The proofs only cover static, single-threaded RISC-V binaries linked against musl libc — but most of us ship x86 or ARM on glibc, where they don't directly apply. And “no-crash” isn't “correct”: AddressSanitizer has blind spots, and pure logic bugs still slip through.
It already caught a decade-old bug
Chasing the proofs surfaced hundreds of real bugs: a rediscovery of CVE-2026-56132 in libexpat, an out-of-bounds write that fires only on a rare 20MB input, and a logic bug that had been latent for over ten years in a widely used application.
Vercel AI SDK 5.0.281 hard-stops rogue tool calls
With toolChoice: 'none', generateText and streamText now throw a ToolChoiceViolationError before any tool executes — so a model that tries to call a tool you deliberately disabled fails fast instead of silently running it.