Artificial Intelligence
Anthropic says Claude proved code correct
Anthropic said Claude, working with Theorem and Rocq, generated formal proofs for existing software in 123 open-source libraries, found bugs and was limited by strict technical conditions.
- Anthropic said Claude, working with Theorem, mathematically verified existing software code rather than writing new code.
- The company said the system produced about 180 million lines of proof code for 123 open-source libraries in a few weeks.
- Anthropic said Claude found hundreds of bugs and in some cases fixed them itself.
- The article said the proofs were checked by Rocq and applied only under narrow conditions, including a specific processor architecture, single-threaded programs and inputs under 64 MB.
Anthropic said its Frontier Red Team, working with Theorem, used Claude to verify code drawn from test programmes in Google’s OSS-Fuzz security project. The company said Rocq checked the proofs that Claude produced.
Anthropic said the model produced about 180 million lines of mathematical proof code in a few weeks. It also said the process needed almost no human review, although Theorem had built the tools and formal foundations.
The company said Claude encountered hundreds of bugs while working and was allowed to repair them. It gave examples involving a PDF reader and another application, and said some faults only appeared under unusual conditions.
Anthropic said formal verification still had strict limits, including one processor architecture, single-threaded programs and inputs below 64 MB. It also said the method had previously been costly, citing seL4 and Airbus, and that Theorem planned further work on glibc, OpenSSL and cURL.
Named in this story
Companies
- Anthropic
- said Claude verified existing software code
- Theorem
- partnered with Anthropic and built the formal tools
- Airbus
- was cited as an example of avionics software using formal verification
Organisations
- Anthropic’s Frontier Red Team
- worked on the proof generation project
- Google’s OSS-Fuzz security project
- supplied test programs from open-source libraries
Products and systems
- Claude
- generated the proof work and fixed bugs
- Rocq
- checked the generated proofs automatically
- seL4
- was cited as an example of costly formal verification
- Opus 5.5
- was said to solve half the proofs with fewer tokens
- Opus 5
- served as the earlier model in that comparison
- glibc
- was listed as a future target for formal verification
- OpenSSL
- was listed as a future target for formal verification
- cURL
- was listed as a future target for formal verification
How the source tells it
The piece was written as a product showcase, with promotional language, roadmap talk and broad extrapolation from a research demo.
- vendor boosterism through repeated promotion of Anthropic’s own claims and roadmap
- speculation presented as fact by turning a limited verification demo into a wider prediction about software automation
- triumph by stressing bug finds and autonomous proof generation as a breakthrough