Formally verify software at large scale with Claude
Anthropic's Frontier Red Team used Claude to formally verify software, proving code is free of certain bugs on codebases larger than was historically feasible.
Done withClaude
- What they did
- The team turned existing OSS-Fuzz fuzz tests into specs: the compiled RISC-V binary never crashes on any input under 64 MB. Using Theorem's RISC-V and Linux semantics, Claude wrote Rocq proofs that a proof checker verified. When a proof hit a bug, Claude patched the code, the spec was regenerated, and Claude continued. Patches were penalized and audited for reward hacking.
- How it went
- Over a few weeks Claude proved no-crash for 239 entry points across 123 libraries, about 180 million lines of Rocq. It also found hundreds of bugs. The proofs cover only single-threaded static RISC-V binaries.
- Worth knowing
- Proofs can take hundreds of millions of tokens each. Agents that could copy each other's proofs used 17% fewer tokens in the 36-task test.
Try it yourself with Claude
I want to formally verify that [a specific module or library in my codebase at path] is free of [a bug class, e.g. out-of-bounds access or integer overflow]. Pick a suitable verification tool for [language], write the specifications and proofs, and run them. Finish with a report listing what was proven, what failed, and any bugs found, and don't change my source code without showing me the diff first.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Discussion on Bluesky · Oct 22, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Generate a daily personalized trading research report from brokerage data
Lonnie VanZandt's Claude-based assistant reads his Schwab positions and watchlist via MCP, sweeps news and Fed sources, and writes a daily report.
Log meals and medical metrics into a health spreadsheet
A Claude Cowork skill tracks diet and medical metrics in a spreadsheet, and meals told to Claude on mobile are logged in the daily record.
Find a Linux vulnerability hidden for 23 years
An account of Claude Code uncovering a long-standing vulnerability in Linux.
Reverse engineer Claude Desktop to find a path to Linux support
A maintainer pointed Claude at the minified Claude Desktop app to find why Cowork only runs on macOS; it found the VM and sandbox layers and mapped a route to Linux.
Solve open math problems autonomously
DeepMind's Aletheia agent on Gemini Deep Think was evaluated on 700 open Erdős problems and autonomously solved four.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.