Google tried this with the original variant of Chronicle and failed.

Threat analysis does not play well with common types of AI/ML applications because the problem space changes so frequently with the design of avoiding predictability. You can gain "awareness" of events using AI/ML (e.g., this road suddenly has tanks moving across it), but it's tough to call that a "threat" (maybe other intel told the analyst those tanks were supposed to be there). Compound that problem to "the world" or even "an entire country" and you run out of compute real quick.

Most solutions err on the side of just popping up alerts but don't identify "threats" until an analyst looks at them.