What are the risks of AI agents and MCP servers, and which controls and tools address them?

Agent security is different from chatbot security: the risk is what the agent does, not just what it says. The core threats are indirect prompt injection through tool results, over-permissioned agents, tool misuse, and malicious or compromised MCP servers. A defense stack in four layers: 1. Discovery and posture: inventory every agent and MCP connection — Zenity covers agents and low-code copilots, Straiker adds agent discovery and posture for agentic apps. 2. Artifact scanning: scan agent code and MCP servers for unsafe permissions and hidden injection paths before deployment (Prisma AIRS 3.0 added exactly this in March 2026). 3. Runtime guardrails: enforce tool-call-level policy so an injected instruction can't trigger unauthorized actions. 4. Continuous red teaming: test agents under tool-misuse scenarios, not just prompt attacks. Benchmark your controls against OWASP's agentic AI security guidance — its tool-poisoning and multi-agent failure controls are starting to appear in enterprise procurement requirements.