AI
Anthropic says Claude generated formal proofs for security properties in 123 libraries
Anthropic and Theorem report using Claude to generate formal proofs for security properties across 123 open-source libraries, covering 239 fuzz-test entry points and about 180 million lines of Rocq proof code. The work also found bugs that conventional fuzzing might miss, including flaws requiring unusually large or structured inputs. The proofs establish bounded properties under specific assumptions—single-threaded static RISC-V binaries, modeled Linux behavior, and inputs below 64 MB—not that the libraries are bug-free. Anthropic says the proofs and trusted components are being published for review.
Editorial summary · Publisher article
What matters
- Claude proved selected no-crash properties after finding and sometimes patching bugs across 239 entry points in 123 libraries.
- Reported bugs included out-of-bounds accesses, a rediscovered libexpat vulnerability, and logic errors that did not cause crashes.
- The proof checker reduces the trusted base to the modeled machine semantics, theorem statements, and checker, according to the report.
- The work does not cover all entry points, all memory-safety bugs, caller misuse, or most deployed x86 and ARM configurations.
Why the result is notable
Formal verification has traditionally required much more specialist effort than testing. The reported workflow applies LLM-generated proofs to compiled machine code and reuses common proof components, making larger-scale verification more plausible.
What remains outside the guarantee
The proofs target defined specifications, not reality in every deployment. They assume a particular CPU, operating-system model, binary format, memory budget, and input bound, while the source acknowledges gaps in specifications, architectures, libraries, entry-point coverage, and caller behavior.
Context & caveats
- These are claims from Anthropic and Theorem; the packet says artifacts are being published for independent verification.
- A formally verified result is only as broad as its specification and machine model.
Why this made the edition
The report contains extensive technical evidence, concrete scale and methodology, published artifacts, examples of discovered bugs, and unusually clear limitations. The claims come from Anthropic and its collaborator, but the described proof-checking and independent-verification caveats are explicit.
Source material