Anthropic Develops Method Using AI to Mathematically Prove Software Is Bug-Free, Aiming for Bug-Free Software

What happened

Anthropic shared a new method for "formal verification." This means writing software code in math-like language and using math to prove it has no bugs. Anthropic worked with the research group Theorem and used its AI model, Claude, to do the work. They tested 123 open-source software libraries, which are free code packages that others can reuse. Each time the proof found a bug, Claude helped fix it and the work continued.

Key facts

• The proof covered code running on a RISC-V processor (a type of CPU chip) with Linux, in a fixed amount of memory, as a single-threaded static binary (one program running in one stream of work).
• Under these conditions, the team proved that bugs causing crashes do not happen.
• The process used more than 100 million tokens (units of text that AI models read and write).
• Anthropic says Claude Opus 5.5 got results with fewer tokens than Claude Opus 5, in a test of 36 verification runs.

Why proofs were rare

The idea of formal verification is not new. But proofs take a lot of human effort, so they were mostly used for life-critical systems, such as the computer code in airplanes. Using AI may make this method practical for much more software.

Why it matters

AI is finding many bugs in software, and some bugs have caused big problems, like an emergency update for PHP. If AI can also help prove that code is safe, software could become more reliable. Anthropic also says the high token cost may drop over time, because better models need fewer tokens and proofs can be reused for different software.

What to watch

• The proof only covers RISC-V. Anthropic says Arm and x86 chips could still have bugs, so more testing is needed.
• Proofs must cover every machine setup where the software runs. Anthropic says this limit can be overcome as methods and models improve.
• Theorem offers its LLM-based verification method free to open-source projects.
• The verification skill used in the work is also public on GitHub.