Prompt-injection defenses for LLM agents — fetch interception, output sanitization, session inspection.