Agents now write most of the boilerplate, but the subtle bugs - an auth check skipped, a query unsanitized - look exactly like correct code. Human review is the bottleneck again. Are there review workflows or tools that actually catch AI-introduced vulnerabilities at scale?