security machine-learning llm prompt-injection indirect-prompt-injection attack llm-agent isolation benchmark defense attention structured-query llm-agents tool-use evaluation ai nlp transformer sequence-to-sequence planning ai-agent detection preference-optimization dpo memory-defense adversarial-ml in-context-learning memory-poisoning authorization tool-call mcp agentic-ai distributed-systems provenance agentic-workflow hpc hallucination responsible-ai