Anthropic Says Claude Proved Security Properties for 239 Entry Points Across 123 Libraries

Anthropic published research on October 22, 2026, reporting that Claude produced machine-checked formal proofs of security properties at the RISC-V assembly code level for 239 fuzz-test entry points across 123 open-source libraries. The work was done in a collaboration with Theorem, according to Anthropic. The details below come from a summary carried by Unite.AI.

What Anthropic reports

• Scope: 239 fuzz-test entry points across 123 open-source libraries.
• Proof level: machine-checked formal proofs at the RISC-V assembly code level.
• Bug fixes: according to Anthropic, Claude patched the bugs it found.
• Property proved: after patching, Claude proved the programs will always exit normally, without triggering memory-safety errors or failing assertions.
• Input bound: the guarantee holds for every input under 64 MB, according to Anthropic.
• Partner: Theorem.

The summary does not name the Claude model used, and no benchmark scores, pricing or context window figures are given for this research.

What this means

Fuzz testing samples inputs and can only show that the tested inputs did not cause a failure. A machine-checked proof of the kind Anthropic describes is a different claim: that the property holds for every input within the stated bound, here 64 MB. Anthropic also says the proofs sit at the RISC-V assembly level, not at the source level. The bug patching and the proving are reported as part of one workflow, with Claude doing both. These are Anthropic's own claims, and the 239 entry points and 123 libraries are the figures that define the scope of what it says was proved.