Anthropic would have found a different library or rolled their own, rather than taking that risk. If the library was fundamental, maybe they'd go for a commercial license, but that's usually an option of last resort.
If you have nearly limitless compute to throw at an issue and a good enough model, then it should be able to create enough test cases to cover most aspects of the codebase (iterating thousands of times until it gets it right) and then eventually write a new implementation in a new language or a slightly different tech stack that passes all of the original tests, alongside a few more hundreds of iterations of refactoring.
I give it a decade until large orgs are doing that to avoid licensing restrictions and other liabilities.