Practical defensive prompt injection: canaries, counter-instructions, red-teaming, and logging that make AI attacks harder to exploit and easier to detect.