You really think AI can't convert unsafe blocks to safe code?
The reason AI can't fix memory unsafety in the original language is the same reason humans can't do it. There isn't enogh feedback from the compiler.
I imagine it could, eventually, if not now. The Bun translation used Anthropic's best model, cost hundreds of thousands of dollars, and still contained a large number of unsafe blocks. Maybe if they left the agent running for longer, using Miri for verification?