Navigating Security Tradeoffs of AI Agents

This report from Unit 42 warns of emerging risks from agentic AI, highlighting three primary threat pathways: malicious model-file uploads that execute payloads when loaded, compromised Model Context Protocol (MCP) servers that can 'rug pull' tool capabilities, and internally deployed AI agents that, if compromised or misconfigured, can perform large-scale fraud, data exfiltration, or unauthorized actions. It recommends defenses including scanning and sandboxing models, using trusted remote MCP providers or auditing local MCP code, implementing prompt-injection guardrails, minimizing agent permissions and tools, detailed logging and provenance, and organizational governance to balance productivity with security.