Oriel — AI & technology worth noticing, every day.

Formally verify software at large scale with Claude

Web page · https://agentusecases.ai/use-case/formally-verify-software-at-large-scale-with-claude

Done 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.

First seen by Oriel on .

AI summary

Anthropic's Frontier Red Team used Claude to produce machine-checked Rocq proofs that large sets of software entry points never crash on inputs under 64 MB.

  • The team converted existing OSS-Fuzz tests into specifications and used Theorem's RISC-V and Linux semantics to generate proofs checked by a proof checker.
  • Over a few weeks, Claude proved 239 entry points across 123 libraries, producing about 180 million lines of Rocq and finding hundreds of bugs.
  • When proofs exposed bugs, Claude patched the code and continued after regenerating the specification; patches were penalized and audited for reward hacking.
  • The results cover only single-threaded static RISC-V binaries, and individual proofs can consume hundreds of millions of tokens.
  • In a 36-task test, agents able to copy each other's proofs used 17% fewer tokens.

#formal verification #claude #rocq #risc-v

Generated by gpt-5.6-sol. Check the original for details.

Aliases

SourceExternal IDURL
Agent Use Caseshttps://agentusecases.ai/use-case/formally-verify-software-at-large-scale-with-claudehttps://agentusecases.ai/use-case/formally-verify-software-at-large-scale-with-claude Read discussion