The whole dayThu 22 Oct 2026
AI Compass
ArchiveAbout

ResearchAnthropicConfirmed

Story 7 of 11 in the Thursday 22 October edition

Anthropic uses Claude to formally verify 239 fuzz-testing targets across 123 open-source libraries

Working with verification startup Theorem, Anthropic’s Frontier Red Team tasked Claude with fixing bugs and generating 180 million lines of machine-checked Rocq proofs directly over compiled RISC-V Linux binaries.

Edited by Ashutosh Shrivastava
22 Oct 2026, 23:55 IST · 2 min read

Split illustration showing classical geometric circle proofs on an olive background on the left and a grid of grey dots with one bright green dot on a coral background on the right
Image: Anthropic

Anthropic’s Frontier Red Team and verification startup Theorem (theorem.dev) have published a large-scale study showing that Claude can autonomously patch memory-safety bugs and write machine-checked formal proofs in the Rocq (formerly Coq) theorem prover for unmodified C, C++ and Rust libraries drawn from Google’s OSS-Fuzz benchmark.

Step chart of share of 36 verification tasks completed against token budget per agent from 0 to 5M tokens comparing Opus 5.5, Fable 5.1 and Opus 5
On a 36-task verification benchmark, Claude Opus 5.5 reached 50% completion using roughly 40% fewer tokens than Claude Opus 5.Image: Anthropic
Stacked horizontal bars comparing identical, lightly edited, edited more and new proof files across libcups, freerdp and libzmq
Sharing a common workspace let freerdp and libzmq reuse allocator and runtime proofs from libcups, cutting token use by 17%.Image: Anthropic

Binary-level Rocq proofs

Rather than asking engineers to rewrite production code in a verification-specific language, the team compiled each OSS-Fuzz target into a standalone RISC-V Linux executable with AddressSanitizer enabled (median size 18,000 instructions) and loaded its machine code into Rocq alongside a formal RISC-V instruction-set specification. Claude agents working inside Claude Code fixed underlying bugs in the source code and then constructed inductive invariants proving that the compiled binary never reaches an invalid instruction, failed assertion or AddressSanitizer memory error on any byte sequence up to 64MB.

Across 239 entry points in 123 libraries, Claude produced roughly 180 million lines of Rocq (~600k lines per proof), published at theorem-labs/oss-verify-proofs.

Functional correctness milestones

Beyond crash-freedom, Claude verified functional correctness on two self-contained systems: it proved in 20 days (~$20k in tokens) that the Linux kernel’s x86 zlib_inflate implementation conforms to RFC 1950 and returns Z_DATA_ERROR on invalid streams, and it wrote a 19,000-line Rust Python interpreter verified in Verus that matched CPython on 99.95% of 1.7M random test programs.

Sources

  1. PrimaryAnthropic Research: Scaling formal verification to real-world softwareanthropic.com
  2. OfficialGitHub: theorem-labs/oss-verify-proofsgithub.com/theorem-labs
  3. On XAnthropic (@AnthropicAI) on Xx.com

The rest of the day

Read the whole day

02Coding agentGoogle

Google releases open-source Gemini API CLI built for coding agents

Published on the Google AI Studio blog on 19 October 2026 and shared on 22 October 2026, the single-binary Rust tool gives coding agents and developers progressive terminal access to Gemini text, image, video, audio, embedding and research APIs without MCP schema overhead.

Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X