Anthropic used its Claude AI models to study 25 open-source software projects. These projects let people see their code. Claude found 493 problems with parts of the code that people could reach without needing them. One problem stayed in the code for 21 years. Two settings worked badly together. A server then let anyone connect, not just trusted users. Claude also found code in OpenSSH that people could reach before logging in. OpenSSH helps people connect safely to other computers. Its developers released fixes. This work does not remove every hidden problem. It makes some problems harder to reach.
In a separate project, Anthropic worked with the software company Theorem. Claude used math to check safety rules in 123 collections of code. These collections help other programs work. It produced proofs for 239 places where programs start work. The proofs showed that these parts would not crash under certain conditions. The work also found hundreds of bugs, or mistakes in code. Some had escaped earlier tests for years. But these proofs do not mean all the software is free of mistakes. They cover only the rules and conditions that the researchers checked.
