Anthropic announced that its Claude model can automatically produce formal verification proofs for existing code. In a test with 123 libraries from Google’s OSS‑Fuzz project, Claude generated roughly 180 million lines of proof in a few weeks. The proofs were checked by the Rocq verifier, and Anthropic says the process required almost no human reviewers.
What the system actually did
Claude did not write new programs. Instead it turned crash‑test cases into mathematical statements, analyzed the compiled machine code, and emitted formal proofs that the code meets those statements. When the model encountered a bug, it was allowed to fix the code itself. In a week of checking a popular PDF reader, Claude uncovered as many security bugs as are usually reported over a full year. In another library it found a logical error that had been hidden for more than ten years.
Limits and costs
Anthropic cautions that the proofs only hold for a single‑threaded program running on a specific processor architecture with inputs under 64 MB. The compute needed is high – individual proofs run into hundreds of millions of tokens – though the company did not disclose total spending. A smaller test showed that the newer Opus 5.5 model can complete half the proofs using about 40 % fewer tokens than its predecessor.
Why it matters
If AI can automatically verify code, developers may spend less time hunting bugs and reviewing patches, especially for safety‑critical software. However, the current approach works only under strict hardware and input limits, so most everyday applications still need traditional testing.
