Oriel — AI & technology worth noticing, every day.

LLMs are making formal software verification feasible at a much larger scale

Web page · https://anthropic.com/research/formal-verification

First seen by Oriel on .

AI summary

Anthropic and Theorem used Claude to prove no-crash properties for compiled RISC-V fuzz tests across 123 open-source libraries, showing that LLMs can scale formal verification while uncovering bugs.

  • Claude proved 239 entry points across 123 OSS-Fuzz libraries, generating about 180 million lines of Rocq proof code checked by the Rocq kernel.
  • The proofs cover every input up to 64 MB under assumptions including a single-threaded static binary, fixed memory budget, RISC-V CPU, and Linux OS.
  • The workflow converts existing fuzz tests into specifications, then has Claude patch discovered bugs and prove the patched programs satisfy those specifications.
  • Discovered issues included rare large-input memory errors, a multi-call out-of-bounds read, a rediscovered libexpat vulnerability, and logic bugs that did not crash.
  • Other experiments covered binary-level zlib verification and a Rust Python interpreter that matched CPython behavior in 99.95% of over 1.7 million generated tests.

#formal verification #llm agents #rocq #software security

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

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Points

Score reported by the source.

  • Hacker News: 454, latest observation ; +423 over observed period since

Comments

Comment count reported by the source.

  • Hacker News: 297, latest observation ; +293 over observed period since

Aliases

SourceExternal IDURL
Hacker News50202217https://www.anthropic.com/research/formal-verification Read discussion