Anthropic used its AI tool Claude to check 25 open-source projects. These projects share their computer code publicly. Claude looked for code that people could reach without a good reason. It found 493 problems. One problem involved two settings. Together, they let anyone connect instead of only trusted users. This problem stayed in the code for 21 years.
Claude also checked OpenSSH, a program for connecting to other computers safely. It found about 4,500 lines of code that people could reach before proving their identity. People responsible for OpenSSH released fixes. Anthropic says limiting access can make hidden problems less dangerous. It does not remove those problems.
In a separate project, Anthropic worked with Theorem, a company that checks software. Claude used maths to check 123 software libraries. These are collections of code for other programs to use. It produced proofs for 239 places where a program can start work. The proofs showed these parts would not crash under certain conditions. The work also found hundreds of bugs, or mistakes in code. Some escaped years of testing. Another tool checked the proofs. However, these checks only cover particular rules and conditions. They do not show that every part of a program is safe.
