Huh. That's actually kind've a clever use case. I hadn't considered that. I presume the main obstacle would be the token limit of whatever LLM that one is using (presuming that it was an LLM that was used). Analyzing an entire codebase, ofc, depending on the project, would likely require an enormous amount of tokens that an LLM wouldn't be able to handle, or it would just be prohibitively expensive. To be clear, that's not to say that I know that such an LLM doesn't exist — one very well could — but if one doesn't, then that would be rationale that i would currently stand behind.

You would first need to define malicious code within the context of that repo. To some people, telemetry is malicious.

Probably not. Obfuscation works, and might even depend on remote code being downloaded at either build time or run time.

There are a lot of heuristics you can use (e.g. disallowing some functions/modules) to check a codebase, but those already exist no AI required. Unless you call static analysis "AI", who knows.