Anthropic recently worked with Theorem to use Claude to formally verify specific security properties across 123 open-source libraries, covering 239 test entry points.
The interesting part is how this differs from ordinary testing.
Testing checks software against selected inputs to find failures. Formal verification uses mathematical proofs to establish that a specific property holds across a defined range of inputs.
In this project, Claude produced proofs showing that the tested programs would not crash for inputs under 64 MB, within a modeled Linux environment and under specific assumptions.
During this process, Claude also found and helped fix hundreds of bugs, including some that conventional fuzz testing was unlikely to catch.
Historically, formal verification has required substantial specialist effort. Anthropic notes that formally verifying the seL4 operating system kernel took about 20 person-years, compared with two person-years to implement it.
Claude completed the large-scale verification effort over a few weeks. However, the process still required hundreds of millions of tokens, making cost and efficiency important challenges.
There is also a limitation worth understanding: this does not mean all 123 libraries are bug-free. The proofs cover specific properties and code paths under defined conditions, not every possible flaw in the software.
What makes this development interesting is that AI could make mathematical verification practical for far more software projects than before.
Instead of relying only on testing to discover failures, developers can increasingly use AI to help establish specific guarantees about how their code behaves.
Research: https://www.anthropic.com/research/formal-verification