From prompt injections to protocol exploits: Threats in LLM-powered AI agents workflows