Anthropic has used its Claude model to create mathematical proofs of the safety of existing code. According to The Decoder, the effort produced around 180 million lines of proof code for 123 open-source libraries. In the process, the model came across bugs that years of testing had not uncovered.
What formal verification means
Classic tests check whether software delivers the expected result in selected situations. However, they only cover cases that someone thought of. Formal verification, by contrast, mathematically proves that a program satisfies a certain property, and it does so for all possible inputs rather than just a sample. The proof itself is also code, which a checking program verifies step by step.
The catch: writing such proofs is laborious and requires specialist knowledge. For this reason, the method has so far mostly been used where errors would be especially costly. This is exactly where the approach comes in. If an AI generates the proofs, the effort could drop significantly.
- TGR-EvoDAG: Agents Evolve Code for Recommenders
- AI Finds Critical PHP Security Vulnerability
- DeepInfer: Deepseek Open-Sources Its Inference Software
Why this matters for developers
According to the source, Anthropic wants to make formal verification via AI as practical as today's standard tests soon. That would be another step in the automation of software development: the AI would not only write code but also secure its correctness.
After the brief report, it remains unclear how much human oversight is needed and which kinds of errors the proofs cover. A proof is only as meaningful as the property it demonstrates. Anyone who uses open-source libraries can keep an eye on the development. However, you should only rely on it once more detailed information is available.